Completely Bounded Maps and Operator Algebras

This page intentionally left blank CAMBRIDGE STUDIES IN ADVANCED MATHEMATICS 78 EDITORIAL BOARD B. BOLLOBAS, W. FULTO...

Author: Vern Paulsen

15 downloads 871 Views 2MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

This page intentionally left blank

CAMBRIDGE STUDIES IN ADVANCED MATHEMATICS 78 EDITORIAL BOARD B. BOLLOBAS, W. FULTON, A. KATOK, F. KIRWAN, P. SARNAK Completely Bounded Maps and Operator Algebras In this book the reader is provided with a tour of the principal results and ideas in the theories of completely positive maps, completely bounded maps, dilation theory, operator spaces, and operator algebras, together with some of their main applications. The author assumes only that the reader has a basic background in functional analysis and C ∗ -algebras, and the presentation is self-contained and paced appropriately for graduate students new to the subject. The book could be used as a text for a course or for independent reading; with this in mind, many exercises are included. Experts will also want this book for their library, since the author presents new and simpler proofs of some of the major results in the area, and many applications are also included. This will be an indispensable introduction to the theory of operator spaces for all who want to know more.

Already published 2 3 5 7 8 9 10 11 12 14 15 16 17 18 19 20 21 22 24 25 26 27 28 29 30 31 32 33 34 35 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73

K. Petersen Ergodic theory P.T. Johnstone Stone spaces J.-P. Kahane Some random series of functions, 2nd edition J. Lambek & P.J. Scott Introduction to higher-order categorical logic H. Matsumura Commutative ring theory C.B. Thomas Characteristic classes and the cohomology of ﬁnite groups M. Aschbacher Finite group theory J.L. Alperin Local representation theory P. Koosis The logarithmic integral I S.J. Patterson An introduction to the theory of the Riemann zeta-function H.J. Baues Algebraic homotopy V.S. Varadarajan Introduction to harmonic analysis on semisimple Lie groups W. Dicks & M. Dunwoody Groups acting on graphs L.J. Corwin & F.P. Greenleaf Representations of nilpotent Lie groups and their applications R. Fritsch & R. Piccinini Cellular structures in topology H. Klingen Introductory lectures on Siegel modular forms P. Koosis The logarithmic integral II M.J. Collins Representations and characters of ﬁnite groups H. Kunita Stochastic ﬂows and stochastic differential equations P. Wojtaszczyk Banach spaces for analysts J.E. Gilbert & M.A.M. Murray Clifford algebras and Dirac operators in harmonic analysis A. Frohlich & M.J. Taylor Algebraic number theory K. Goebel & W.A. Kirk Topics in metric ﬁxed point theory J.F. Humphreys Reﬂection groups and Coxeter groups D.J. Benson Representations and cohomology I D.J. Benson Representations and cohomology II C. Allday & V. Puppe Cohomological methods in transformation groups C. Soule et al. Lectures on Arakelov geometry A. Ambrosetti & G. Prodi A primer of nonlinear analysis J. Palis & F. Takens Hyperbolicity, stability and chaos at homoclinic bifurcations Y. Meyer Wavelets and operators 1 C. Weibel An introduction to homological algebra W. Bruns & J. Herzog Cohen-Macaulay rings V. Snaith Explicit Brauer induction G. Laumon Cohomology of Drinfeld modular varieties I E.B. Davies Spectral theory and differential operators J. Diestel, H. Jarchow, & A. Tonge Absolutely summing operators P. Mattila Geometry of sets and measures in Euclidean spaces R. Pinsky Positive harmonic functions and diffusion G. Tenenbaum Introduction to analytic and probabilistic number theory C. Peskine An algebraic introduction to complex projective geometry Y. Meyer & R. Coifman Wavelets R. Stanley Enumerative combinatorics I I. Porteous Clifford algebras and the classical groups M. Audin Spinning tops V. Jurdjevic Geometric control theory H. Volklein Groups as Galois groups J. Le Potier Lectures on vector bundles D. Bump Automorphic forms and representations G. Laumon Cohomology of Drinfeld modular varieties II D. M. Clark & B. A. Davey Natural dualities for the working algebraist J. McCleary A user’s guide to spectral sequences II P. Taylor Practical foundations of mathematics M.P. Brodmann & R.Y. Sharp Local cohomology J.D. Dixon et al. Analytic pro-P groups R. Stanley Enumerative combinatorics II R. M. Dudley Uniform central limit theorems J. Jost & X. Li-Jost Calculus of variations A.J. Berrick & M.E. Keating An introduction to rings and modules S. Morosawa Holomorphic dynamics A.J. Berrick & M.E. Keating Categories and modules with K-theory in view K. Sato Levy processes and inﬁnitely divisible distributions H. Hida Modular forms and Galois cohomology R. Iorio & V. Iorio Fourier analysis and partial differential equations R. Blei Analysis in integer and fractional dimensions F. Borceaux & G. Janelidze Galois theories B. Bollobas Random graphs

COMPLETELY BOUNDED MAPS AND OPERATOR ALGEBRAS VERN PAULSEN University of Houston

   Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge  , United Kingdom Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521816694 © Vern Paulsen 2002 This book is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format 2003 - isbn-13 978-0-511-06949-9 eBook (EBL) - isbn-10 0-511-06949-9 eBook (EBL) - isbn-13 978-0-521-81669-4 hardback - isbn-10 0-521-81669-6 hardback

Cambridge University Press has no responsibility for the persistence or accuracy of s for external or third-party internet websites referred to in this book, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

To John, Ival, Efﬁe, Susan, Stephen and Lisa. My past, present and future.

Contents

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Preface Introduction Positive Maps Completely Positive Maps Dilation Theorems Commuting Contractions on Hilbert Space Completely Positive Maps into Mn Arveson’s Extension Theorems Completely Bounded Maps Completely Bounded Homomorphisms Polynomially Bounded and Power-Bounded Operators Applications to K -Spectral Sets Tensor Products and Joint Spectral Sets Abstract Characterizations of Operator Systems and Operator Spaces An Operator Space Bestiary Injective Envelopes Abstract Operator Algebras Completely Bounded Multilinear Maps and the Haagerup Tensor Norm Universal Operator Algebras and Factorization Similarity and Factorization Bibliography Index

vii

page ix 1 9 26 43 58 73 84 97 120 135 150 159 175 186 206 225 239 260 273 285 297

Preface

This book is intended to give the reader an introduction to the principal results and ideas in the theories of completely positive maps, completely bounded maps, dilation theory, operator spaces, and operator algebras, together with some of their main applications. It is intended to be self-contained and accessible to any reader who has had a ﬁrst course in functional analysis that included an introduction to C ∗ -algebras. It could be used as a text for a course or for independent reading. With this in mind, we have included plenty of exercises. We have made no attempt at giving a full state-of-the-art exposition of any of these ﬁelds. Instead, we have tried to give the reader an introduction to many of the important techniques and results of these ﬁelds, together with a feel for their connections and some of the important applications of the ideas. However, we present new proofs and approaches to some of the well-known results in this area, which should make this book of interest to the expert in this area as well as to the beginner. The quickest route to a result is often not the most illuminating. Consequently, we occasionally present more than one proof of some results. For example, scattered throughout the text and exercises are ﬁve different proofs of a key inequality of von Neumann. We feel that such redundancy can lead to a deeper understanding of the material. In an effort to establish a common core of knowledge that we can assume the reader is familiar with, we have adopted R.G. Douglas’s Banach Algebra Techniques in Operator Theory as a basic text. Results that appear in that text we have assumed are known, and we have attempted to give a full accounting of all other facts by either presenting them, leaving them as an exercise, or giving a reference. Consequently, parts of the text may seem unnecessarily elementary to some readers. For example, readers with a background in Banach spaces or C ∗ -algebras may ﬁnd our discussions of the tensor theory a bit na¨ıve. We now turn our attention to a description of the contents of this book. ix

x

Preface

The ﬁrst seven chapters develop the theory of positive and completely positive maps together with their connections with dilation theory. Dilations are a technique for studying operators on a Hilbert space by representing a given operator as the restriction of a (hopefully) better-understood operator, acting on a larger Hilbert space, to the original space. The operator on the larger space is referred to as a dilation of the original operator. Thus, dilation theory involves essentially geometric constructions. We shall see that many of the classic theorems that characterize which sequences of complex numbers are the moments of a measure are really dilation theorems. One of the better-known dilation theorems is due to Sz.-Nagy and asserts that every contraction operator can be dilated to a unitary operator. Thus to prove some results about contraction operators it is enough to show that they are true for unitary operators. The most famous application of this idea is Sz.-Nagy’s elegant proof of an inequality of von Neumann to the effect that the norm of a polynomial in a contraction operator is at most the supremum of the absolute value of the polynomial over the unit disk. Ando generalized Sz.-Nagy’s and von Neumann’s results to pairs of commuting contractions, but various analogues of these theorems are known to fail for three or more commuting contractions. Work of Sarason and of Sz.-Nagy and Foias showed that many classical results about analytic functions, including the Nevanlinna–Pick theory, Nehari’s theorem, and Caratheodory’s completion theorem are consequences of these results about contraction operators. Thus, one ﬁnds that there is an operator-theoretic obstruction to generalizing many of these classic results. These results are the focus of Chapter 5. W.F. Stinespring introduced the theory of completely positive maps as a means of giving abstract necessary and sufﬁcient conditions for the existence of dilations. In many ways completely positive maps play the same role as positive measures when commutative C ∗ -algebras are replaced by noncommutative C ∗ -algebras. The connections between completely positive maps and dilation theory were broadened further by Arveson, who developed a deep structure theory for these maps, including an operator-valued Hahn–Banach extension theorem. Completely positive maps also play a central role in the theory of tensor products of C ∗ -algebras. Characterizations of nuclear C ∗ -algebras and injectivity are given in terms of these maps. In noncommutative harmonic analysis they arise in the guise of positive deﬁnite operator-valued functions on groups. In spite of the broad range of applications of completely positive maps, this text is one of the few places where one can ﬁnd a full introduction to their theory. In the early 1980s, motivated largely by the work of Wittstock and Haagerup, researchers began extending much of the theory of completely positive maps to

Preface

xi

the family of completely bounded maps. To the extent that completely positive maps are the analogue of positive measures, completely bounded maps are the analogue of bounded measures. This newer family of maps also allows for the development of a theory that ties together many questions about the existence or nonexistence of similarities or what are sometimes referred to as skew dilations. Two famous problems of this type are Kadison’s and Halmos’s similarity conjectures. The theory of completely bounded maps has had an enormous impact on both of these conjectures. Kadison conjectured that every bounded homomorphism of a C ∗ -algebra into the algebra of operators on a Hilbert space is similar to a ∗-homomorphism. Halmos conjectured that every polynomially bounded operator is similar to a contraction. In Chapters 8 and 9, we develop the basic theory of completely bounded maps and their connections with similarity questions. In Chapter 10 we study polynomially bounded operators and present Pisier’s counterexample to the Halmos conjecture. The Kadison conjecture still remains unresolved at the time of this writing, but in Chapter 19 we present Pisier’s theory of similarity and factorization degrees, which we believe is the most hopeful route towards a solution of the Kadison conjecture. Attempts to generalize von Neumann’s results and the theory of polynomially bounded operators to domains other than the unit disk led to the concepts of spectral and K -spectral sets. We study the applications of the theory of completely bounded maps to these ideas in Chapter 11. In Chapter 12 we get our ﬁrst introduction to tensor theory in order to further develop some of the multivariable analogues of von Neumann’s inequality. In order to discuss completely positive or completely bounded maps between two spaces, the domains and ranges of these maps need to be what is known as an operator system or operator space, respectively. Such spaces arise naturally as subspaces of the space of bounded operators on a Hilbert space, and this is how operator systems and operator spaces were originally deﬁned. However, results of Choi and Effros and of Ruan gave abstract characterizations of operator systems and operator spaces that enabled researchers to treat their theory and the corresponding theories of completely positive and completely bounded maps in a way that was free of dependence on this underlying Hilbert space. These abstract characterizations have had an impact on this ﬁeld similar to the impact that the Gelfand–Naimark–Segal theorem has had on the study of C ∗ -algebras. These characterizations have also allowed for the development of many parallels with ideas from the theory of Banach spaces and bounded linear maps, which have in turn led to a deeper understanding of many results in the theory of C ∗ -algebras and von Neumann algebras.

xii

Preface

We present these characterizations in Chapter 13 and give a quick introduction to the rapidly growing ﬁeld of operator spaces in Chapter 14, where we carefully examine some of the more important examples of operator spaces. The abstract characterization of operator spaces led to the Blecher–Ruan– Sinclair abstract characterization of operator algebras. The last chapters of this book are devoted to a development of this theory and some of its applications. We give two separate developments of the Blecher–Ruan–Sinclair theory. First, we present a new proof based on Hamana’s theory of injective envelopes, which we develop in Chapters 15 and 16. We then develop the theory of the Haagerup tensor product and the representation theorems for multilinear maps that are completely bounded in the sense of Christensen and Sinclair, and give a proof of the Blecher–Ruan–Sinclair theorem based on this theory. Our development of the Haagerup tensor theory also uses the theory of the injective envelope in a novel fashion. The remaining two chapters of the book develop some applications of the Blecher–Ruan–Sinclair theorem. First, we develop the theory of the universal operator algebra of a unital algebra and its applications, including new proofs of Nevanlinna’s factorization theorem for analytic functions on the disk and Agler’s generalization of Nevanlinna’s theorem to analytic functions on the bidisk. Finally, in the last chapter, we present Pisier’s theory of the universal operator algebra of an operator space, his results on similarity and factorization degree, and their applications to Kadison’s similarity conjecture. This book grew out of my earlier lecture notes on completely positive and completely bounded maps [161], that have been out-of-print for over a decade. I would like to acknowledge all the friends and colleagues who have helped to make this book possible. David Blecher, Ken Davidson, Gilles Pisier, and Roger Smith have been tremendous sources of information and ideas. The reader should be grateful to Aristides Katavolos, whose proofreading of selected chapters led to a further polishing of the entire manuscript. Ron Douglas has been a constant source of support throughout my academic career and provided the original impetus to write this book. The author would also like to thank the Mathematics Department at Rice University where portions of this book were written, Roger Astley of Cambridge University Press for his support and advice, the proofreaders at TechBooks for making many of my thoughts ﬂow smoother without altering their mathematical content, and Robin Campbell who typed nearly the entire manuscript and contributed much to its overall look. Finally, without my family’s patience and endurance this project would have not been possible. While writing this book I was partially supported by a grant from the National Science Foundation.

Chapter 1 Introduction

It is assumed throughout this book that the reader is familiar with operator theory and the basic properties of C ∗ -algebras (see for example [76] and [8, Chapter 1]). We concentrate primarily on giving a self-contained exposition of the theory of completely positive and completely bounded maps between C ∗ -algebras and the applications of these maps to the study of operator algebras, similarity questions, and dilation theory. In particular, we assume that the reader is familiar with the material necessary for the Gelfand–Naimark–Segal theorem, which states that every C ∗ -algebra has a one-to-one, ∗-preserving, norm-preserving representation as a norm-closed, ∗-closed algebra of operators on a Hilbert space. In this chapter we introduce some of the key concepts that will be studied in this book. As well as having a norm, a C ∗ -algebra also has an order structure, induced by the cone of positive elements. Recall that an element of a C ∗ -algebra is positive if and only if it is self-adjoint and its spectrum is contained in the nonnegative reals, or equivalently, if it is of the form a ∗ a for some element a. Since the property of being positive is preserved by ∗-isomorphism, if a C ∗ algebra is represented as an algebra of operators on a Hilbert space, then the positive elements of the C ∗ -algebra coincide with the positive operators that are contained in the representation of the algebra. An equivalent characterization of positivity for an operator on a Hilbert space is that A is a positive operator provided that the inner product Ax, x is nonnegative for every vector x in the space. We shall write a ≥ 0 to denote that a is positive. The positive elements in a C ∗ -algebra A are a norm-closed, convex cone in the C ∗ -algebra, denoted by A+ . If h is a self-adjoint element, then it is easy to see, via the functional calculus, that h is the difference of two positive elements.

1

2

Chapter 1. Introduction

Indeed, if we let f + (x) =

x, x ≥ 0, 0, x < 0,

f − (x) =

0, x ≥ 0, −x, x < 0,

then using the functional calculus we have that h = f + (h) − f − (h), with f + (h) and f − (h) both positive. In particular, we see that the real linear span of the positive elements is the set of self-adjoint elements, which is also norm-closed. Using the Cartesian decomposition of an arbitrary element a of A, namely, a = h + ik with h = h ∗ , k = k ∗ , we see that a = ( p1 − p2 ) + i( p3 − p4 ), with pi positive, i = 1, 2, 3, 4. Thus, the complex linear span of A+ is A. In addition to having its own norm and order structure, a C ∗ -algebra is also equipped with a whole sequence of norms and order structures on a set of C ∗ -algebras naturally associated with the original algebra, and this additional structure will play a central role in this book. To see how to obtain this additional structure, let A be our C ∗ -algebra, let Mn denote the n × n complex matrices, and let Mn (A) denote the set of n × n matrices with entries from A. We’ll denote a typical element of Mn (A) by (ai, j ). There is a natural way to make Mn (A) into a ∗-algebra. Namely, for (ai, j ) and (bi, j ) in Mn (A), set (ai, j ) · (bi, j ) =

n

ai,k bk, j

k=1

and (ai, j )∗ = (a ∗j,i ). What is not so obvious is that there is a unique way to introduce a norm such that Mn (A) becomes a C ∗ -algebra. To see how this is done, we begin with the most basic of all C ∗ -algebras, B(H), the bounded linear operators on a Hilbert space H. If we let H(n) denote the direct sum of n copies of H, then there is a natural norm and inner product on H(n) that makes it into a Hilbert space. Namely,  2 h1  ..   .  = h 1 2 + · · · + h n 2 h n

Chapter 1. Introduction and

    h1 k1 ,  ..   ..   . ,  .  hn

where

kn

3

= h 1 , k1 H + · · · + h n , kn H , H

(n)

   h1 k1  ..   ..   .  and  .  

hn

kn

are in H(n) . This Hilbert space is also often denoted 2n (H). We prefer to regard elements of H(n) as column vectors, for reasons that will become apparent shortly. There is a natural way to regard an element of Mn (B(H)) as a linear map on (n) H , by using the ordinary rules for matrix products. That is, we set   n T1 j h j     j=1 h1     ..   . .. (Ti j )  .  =  ,  n    hn Tn j h j j=1

h 1 . for (Ti j ) in Mn (B(H)) and .. in H(n) . It is easily checked (Exercise 1.1) hn

that every element of Mn (B(H)) deﬁnes a bounded linear operator on H(n) and that this correspondence yields a one-to-one ∗-isomorphism of Mn (B(H)) onto B(H(n) ) (Exercise 1.2). Thus, the identiﬁcation Mn (B(H)) = B(H(n) ) gives us a norm that makes Mn (B(H)) a C ∗ -algebra. Now, given any C ∗ -algebra A, one way that Mn (A) can be viewed as a C ∗ algebra is to ﬁrst choose a one-to-one ∗-representation of A on some Hilbert space H so that A can be identiﬁed as a C ∗ -subalgebra of B(H). This allows us to identify Mn (A) as a ∗-subalgebra of Mn (B(H)). It is straightforward to verify that the image of Mn (A) under this representation is closed and hence a C ∗ -algebra. Thus, by using a one-to-one *-representation of A, we have a way to turn Mn (A) into a C ∗ -algebra. But since the norm is unique on a C ∗ -algebra, we see that the norm on Mn (A) deﬁned in this fashion is independent of the particular representation of A that we chose. Since positive elements remain positive under ∗-isomorphisms, we see that the positive elements of Mn (A) are also uniquely determined.

4

Chapter 1. Introduction

So we see that in addition to having a norm and an order, every C ∗ -algebra A carries along this extra “baggage” of canonically deﬁned norms and orders on each Mn (A). Remarkably, keeping track of how this extra structure behaves yields far more information than one might expect. The study of these matrix norms and matrix orders will be a central topic of this book. For some examples of this structure, we ﬁrst consider Mk . We can regard this as a C ∗ -algebra by identifying Mk with the linear transformations on kdimensional (complex) Hilbert space, Ck . There is a natural way to identify Mn (Mk ) with Mnk , namely, forget the additional parentheses. It is easy to see that, with this identiﬁcation, the multiplication and ∗-operation on Mn (Mk ) become the usual multiplication and ∗-operation on Mnk , that is, the identiﬁcation deﬁnes a ∗-isomorphism. Hence, the unique norm on Mn (Mk ) is just the norm obtained by this identiﬁcation with Mnk . An element of Mn (Mk ) will be positive if and only if the corresponding matrix in Mnk is positive. For a second example, let X be a compact Hausdorff space, and let C(X ) denote the continuous complex-valued functions on X . Setting f ∗ (x) = f (x), we have

f = sup{| f (x)|: x ∈ X }, and deﬁning the algebra operations pointwise makes C(X ) into a C ∗ -algebra. An element F = ( f i, j ) of Mn (C(X )) can be thought of as a continuous Mn valued function. Note that addition, multiplication, and the ∗-operation in Mn (C(X )) are just the pointwise addition, pointwise multiplication, and pointwise conjugate-transpose operations of these matrix-valued functions. If we set

F = sup{ F(x) : x ∈ X }, where by F(x) we mean the norm in Mn , then it is easily seen that this deﬁnes a C ∗ -norm on Mn (C(X )), and thus is the unique norm in which Mn (C(X )) is a C ∗ -algebra. Note that the positive elements of Mn (C(X )) are those F for which F(x) is a positive matrix for all x. Now, given two C ∗ -algebras A and B and a map φ: A → B, we also obtain maps φn : Mn (A) → Mn (B) via the formula φn ((ai, j )) = (φ(ai, j )). In general the adverb completely means that all of the maps {φn } enjoy some property. For example, the map φ is called positive if it maps positive elements of A to positive elements of B, and φ is called completely positive if every φn is a positive map.

Chapter 1. Introduction

5

In a similar fashion, if φ is a bounded map, then each φn will be bounded, and when φ cb = supn φn is ﬁnite, we call φ a completely bounded map. One’s initial hope is perhaps that C ∗ -algebras are sufﬁciently nice that every positive map is completely positive and every bounded map is completely bounded. Indeed, one might expect that φ = φn for all n. For these reasons, we begin with an example of a fairly nice map where those norms are different. Let {E i, j }i,2 j=1 denote the system of matrix units for M2 [that is, E i, j is 1 in the (i, j)th entry and 0 elsewhere], and let φ: M2 → M2 be the transpose map, so that φ(E i, j ) = E j,i . It is easy to verify (Exercise 1.9) that the transpose of a positive matrix is positive and that the norm of the transpose of a matrix is the same as the norm of the matrix, so φ is positive and φ = 1. Now let’s consider φ2 : M2 (M2 ) → M2 (M2 ). Note that the matrix of matrix units,   1 0 0 1 0 0 0 0 E 11 E 12  = 0 0 0 0, E 21 E 22 1 0 0 1 is positive, but that φ2

E 11 E 12 E 21 E 22

φ(E 11 ) = φ(E 21 )

 1 0 φ(E 12 ) = 0 φ(E 22 ) 0

0 0 1 0

0 1 0 0

 0 0  0 1

is not positive. Thus, φ is a positive map but not completely positive. In a similar fashion, we have that E 11 E 21 = 1, E 12 E 22 while the norm of its image under φ2 has norm 2. Thus, φ2 ≥ 2, so φ2 =

φ . It turns out that φ is completely bounded, in fact, supn φn = 2, as we shall see later in this book. To obtain an example of a map that’s not completely bounded, we need to repeat the above example but on an inﬁnite-dimensional space. So let H be a separable, inﬁnite-dimensional Hilbert space with a countable, orthonormal basis, {en }∞ n=1 . Every bounded, linear operator T on H can be thought of as an inﬁnite matrix whose (i, j)th entry is the inner product T e j , ei . One then deﬁnes a map φ from the C ∗ -algebra of bounded linear operators on H, B(H), to B(H) by the transpose. Again φ will be positive and an isometry (Exercise 1.9), but φn ≥ n. To see this last claim, let {E i, j }i,∞j=1 be matrix units on H, and

6

Chapter 1. Introduction

for ﬁxed n, let A = (E j,i ), that is, A is the element of Mn (B(H)) whose (i, j)th entry is E j,i . We leave it to the reader to verify that A = 1 (in fact, A is a partial isometry), but φn (A) = n (Exercise 1.8). There is an alternative approach to the above constructions, via tensor products. A reader familiar with tensor products has perhaps realized that the algebra Mn (A) that we’ve deﬁned is readily identiﬁed with the tensor product algebra Mn ⊗ A. Recall that one makes the tensor product of two algebras into an algebra by deﬁning (a1 ⊗ b1 ) · (a2 ⊗ b2 ) = (a1 a2 ) ⊗ (b1 b2 ) and then extending linearly. If {E i, j }i,n j=1 denotes the canonical basis for Mn , then an element (ai, j ) in Mn (A) can be identiﬁed with i,n j=1 ai, j ⊗ E i, j in Mn ⊗ A. We leave it to the reader to verify (Exercise 1.10) that with this identiﬁcation of Mn (A) and Mn ⊗ A, the multiplication deﬁned on Mn (A) becomes the tensor product multiplication on Mn ⊗ A. Thus, this identiﬁcation is an isomorphism of these algebras. We shall on occasion return to this tensor product notation to simplify concepts. Now that the reader has been introduced to the concepts of completely positive and completely bounded maps, we turn to the topic of dilations. In general, the key idea behind a dilation is to realize an operator or a mapping into a space of operators as “part” of something simpler on a larger space. The simplest case is the unitary dilation of an isometry. Let V be an isometry on H, and let P = IH − V V ∗ be the projection onto the orthocomplement of the range of V . If we deﬁne U on H ⊕ H = K via V P , U= 0 V∗ then it is easily checked that U ∗ U = UU ∗ = IK , so that U is a unitary on K. Moreover, if we identify H with H ⊕ 0, then V n = PH U n |H for all n ≥ 0. Thus, any isometry V can be realized as the restriction of some unitary to one of its subspaces in a manner that also respects the powers of both operators. In a similar fashion, one can construct an isometric dilation of a contraction. Let T be an operator on H, T ≤ 1, and let DT = (I − T ∗ T )1/2 . Note that

T h 2 + DT h 2 = T ∗ T h, h + DT2 h, h = h 2 . We set ∞ 2 2 (H) = (h 1 , h 2 , . . . ): h n ∈ H for all n,

h n < +∞ . n=1

Chapter 1. Introduction 7 2 This is a Hilbert space with (h 1 , h 2 , . . . ) 2 = ∞ n=1 h n , and inner product ∞ (h 1 , h 2 , . . . ), (k1 , k2 , . . . ) = n=1 h n , kn . We deﬁne V : 2 (H) → 2 (H) via V ((h 1 , h 2 , . . . )) = (T h 1 , DT h 1 , h 2 , . . . ). Since V ((h 1 , h 2 , . . . )) 2 = T h 1 2 + DT h 1 2 + h 2 2 + · · · = (h 1 , h 2 , . . . ) 2 , V is an isometry on 2 (H). If we identify H with H ⊕ 0 ⊕ · · · , then it is clear that T n = PH V n |H for all n ≥ 0. Combining these two constructions yields the unitary dilation of a contraction. Theorem 1.1 (Sz.-Nagy’s dilation theorem). Let T be a contraction operator on a Hilbert space H. Then there is a Hilbert space K containing H as a subspace and a unitary operator U on K such that T n = PH U n |H . Proof. Let K = 2 (H) ⊕ 2 (H), and identify H with (H ⊕ 0 ⊕ · · · ) ⊕ 0. Let V be the isometric dilation of T on 2 (H), and let U be the unitary dilation of V on 2 (H) ⊕ 2 (H). Since H ⊆ 2 (H) ⊕ 0, we have that PH U n |H = PH V n |H = T n for all n ≥ 0. Whenever Y is an operator on a Hilbert space K, H is a subspace of K, and X = PH Y |H , then we call X a compression of Y . There is a certain sense in which a “minimal” unitary dilation can be chosen, and this dilation is in some sense unique. We shall not need these facts now, but shall return to them in Chapter 4. To see the power of this simple geometric construction, we now give Sz.-Nagy’s proof of an inequality due to von Neumann. Corollary 1.2 (von Neumann’s inequality). Let T be a contraction on a Hilbert space. Then for any polynomial p,

p(T ) ≤ sup{| p(z)|: |z| ≤ 1}. Proof. Let U be a unitary dilation of T . Since T n = PH U n |H for all n ≥ 0, it follows, by taking linear combinations, that p(T ) = PH p(U )|H , and hence p(T ) ≤ p(U ) . Since unitaries are normal operators, we have that

p(U ) = sup{| p(λ)|: λ ∈ σ (U )}, where σ (U ) denotes the spectrum of U . Finally, since U is unitary, σ (U ) is contained in the unit circle and the result follows.

8

Chapter 1. Introduction

In Chapter 2, we will give another proof of von Neumann’s inequality, using some facts about positive maps, and then in Chapter 4 we will obtain Sz.-Nagy’s dilation theorem as a consequence of von Neumann’s inequality.

Exercises 1.1 Let (Ti j ) be in Mn (B(H)). Verify that the linear transformation it deﬁnes on H(n) is bounded and that, in fact, (Ti j ) ≤ ( i,n j=1 Ti j 2 )1/2 . 1.2 Let π : Mn (B(H)) → B(H(n) ) be the identiﬁcation given in the text. (i) Verify that π is a one-to-one ∗-homomorphism. (ii) Let E j : H → H(n) be the map deﬁned by setting E j (h) equal to the vector that has h for its jth component and is 0 elsewhere. Show that E ∗j : H(n) → H is the map that sends a vector in H(n) to its jth component. (iii) Given T ∈ B(H(n) ), set Ti j = E i∗ T E j . Show that π ((Ti j )) = T and that consequently π is onto. 1.3 Let (Ti j ) be in Mn (B(H)). Prove that (Ti j ) is a contraction if and only if for every choice of 2n unit vectors x1 , . . . , xn , y1 , . . . , yn in H, the scalar matrix (Ti j x j , yi ) is a contraction. 1.4 Let (Ti j ) be in Mn (B(H)). Prove that (Ti j ) is positive if and only if for every choice of n vectors x1 , . . . , xn in H the scalar matrix (Ti j x j , xi ) is positive. 1.5 Let A and B be unital C ∗ -algebras, and let π : A → B be a ∗-homomorphism with π(1) = 1. Show that π is completely positive and completely bounded and that π = πn = π cb = 1. 1.6 Let A, B, and C be C ∗ -algebras, and let φ: A → B and ψ: B → C be (completely) positive maps. Show that ψ ◦ φ is (completely) positive. 1.7 Let {E i, j }i,n j=1 be matrix units for Mn , let A = (E j,i )i,n j=1 , and let B = (E i, j )i,n j=1 be in Mn (Mn ). Show that A is unitary and that n1 B is a rank one projection. 1.8 Let {E i, j }i,∞j=1 be a system of matrix units for B(H), let A = (E j,i )i,n j=1 , and let B = (E i, j )i,n j=1 be in Mn (B(H)). Show that A is a partial isometry, and that n1 B is a projection. Show that φn (A) = B and φn (A) = n. 1.9 Let A be in Mn , and let At denote the transpose of A. Prove that A is positive if and only if At is positive, and that A = At . Prove that these same results hold for operators on a separable, inﬁnite-dimensional Hilbert space, when we ﬁx an orthonormal basis, regard operators as inﬁnite matrices, and use this to deﬁne a transpose map. 1.10 Prove that the map π: Mn (A) → Mn ⊗ A deﬁned by π ((ai, j )) = n i, j=1 ai, j ⊗ E i, j is an algebra isomorphism.

Chapter 2 Positive Maps

Before turning our attention to the completely positive or completely bounded maps, we begin with some results on positive maps that we shall need repeatedly. These results also serve to illustrate how many simpliﬁcations arise when one passes to the smaller class of completely positive maps. If S is a subset of a C ∗ -algebra A, then we set S ∗ = {a: a ∗ ∈ S}, and we call S self-adjoint when S = S ∗ . If A has a unit 1 and S is a selfadjoint subspace of A containing 1, then we call S an operator system. If S is an operator system and h is a self-adjoint element of S, then even though f + (h) and f − (h) need not belong to S (since these only belong to the normclosed algebra generated by h), we can still write h as the difference of two positive elements in S. Indeed, h=

1 1 (h · 1 + h) − (h · 1 − h). 2 2

If S is an operator system, B is a C ∗ -algebra, and φ: S → B is a linear map, then φ is called a positive map provided that it maps positive elements of S to positive elements of B. In this chapter, we develop some of the properties of positive maps. In particular, we shall be concerned with how the assumption of positivity is related to the norm of the map, and conversely, when assumptions about the norm of a map guarantee that it is positive. We give a fairly elementary proof of von Neumann’s inequality (Corollary 2.7), which only uses these observations about positive maps and an elementary result from complex analysis due to Fejer and Riesz. If φ is a positive, linear functional on an operator system S, then it is easy to show that φ = φ(1) (Exercise 2.3). When the range is a C ∗ -algebra the situation is quite different. 9

10

Chapter 2. Positive Maps

Proposition 2.1. Let S be an operator system, and let B be a C ∗ -algebra. If φ: S → B is a positive map, then φ is bounded and φ ≤ 2φ(1). Proof. First note that if p is positive, then 0 ≤ p ≤ p · 1 and so 0 ≤ φ( p) ≤ p · φ(1), from which it follows that φ( p) ≤ p · φ(1) when p ≥ 0. Next note that if p1 and p2 are positive, then p1 − p2 ≤ max{ p1 , p2 }. If h is self-adjoint in S, then using the above decomposition of h, we have 1 1 φ(h · 1 + h) − φ(h · 1 − h), 2 2 which expresses φ(h) as a difference of two positive elements of B. Thus, φ(h) =

1 max{φ(h · 1 + h), φ(h · 1 − h)} ≤ h · φ(1). 2 Finally, if a is an arbitrary element of S, then a = h + ik with h, k ≤ a, h = h ∗ , k = k ∗ , and so φ(h) ≤

φ(a) ≤ φ(h) + φ(k) ≤ 2a · φ(1).

Let us reproduce an example of Arveson, which shows that 2 is the best constant in Proposition 2.1. Example 2.2. Let T denote the unit circle in the complex plane, C(T) the continuous functions on T, z the coordinate function, and S ⊆ C(T) the subspace spanned by 1, z, and z¯ . We deﬁne φ: S → M2 by a 2b φ(a + bz + c¯z ) = . 2c a We leave it to the reader to verify that an element a1 + bz + c¯z of S is positive if and only if c = b¯ and a ≥ 2|b|. It is fairly standard that a self-adjoint element of M2 is positive if and only if its diagonal entries and its determinant are nonnegative real numbers. Combining these two facts, it is clear that φ is a positive map. However, 2φ(1) = 2 = φ(z) ≤ φ, so that φ = 2φ(1). The existence of unital, positive maps that are not contractive can be roughly attributed to two factors. One is the noncommutativity of the range, the other

Chapter 2. Positive Maps

11

is the lack of sufﬁciently many positive elements in the domain. This ﬁrst principle is illustrated in the exercises, and we concentrate here on properties of the domain that ensure that unital, positive maps are contractive. Lemma 2.3. Let A be a C ∗ -algebra with unit, and let pi , i = 1, . . . , n, be positive elements of A such that n

pi ≤ 1.

i=1

If λi , i = 1, . . . , n, are scalars with |λi | ≤ 1, then n λi pi ≤ 1. i=1 Proof. Note that n    



i=1

0 .. . 0

λi pi 0 · · · 0



1/2

p1

  0 · · · 0  0 .. . . ..  =  .. . .  . . 0 ··· 0 0  λ1  0 ×  .. . 0

...

1/2 

pn

...

 0  ..  . 

...

0

   1/2 0 ... 0 0 ... 0 ..   p1 .. .. .. ..  . . .   .. · . .. .. . . . . . 0 1/2 pn 0 ... 0 . . . 0 λn

n The norm of the matrix on the left is i=1 λi pi , while each of the three matrices on the right can be easily seen to have norm less than 1, by using the fact that a ∗ a = aa ∗ = a2 . Theorem 2.4. Let B be a C ∗ -algebra with unit, let X be a compact Hausdorff space, with C(X ) the continuous functions on X , and let φ: C(X ) → B be a positive map. Then φ = φ(1). Proof. By scaling, we may assume that φ(1) ≤ 1. Let f ∈ C(X ), f ≤ 1, and let ε > 0 be given. First, we note that, by a standard partition-of-unity argument, f may be approximated to within ε by a sum of the form given in n of X such that Lemma 2.3. To see this, ﬁrst choose a ﬁnite open covering {Ui }i=1 | f (x) − f (xi )| < ε for x in Ui , and let { pi } be a partition of unity subordinate

12

Chapter 2. Positive Maps

to the covering. That is, { pi } are nonnegative continuous functions satisfying n / Ui , i = 1, . . . , n. Set λi = f (xi ), and note i=1 pi = 1 and pi (x) = 0 for x ∈ that if pi (x) = 0 for some i, then x ∈ Ui and so | f (x) − λi | < ε. Hence, for any x, λi pi (x) = ( f (x) − λi ) pi (x) f (x) − ≤ | f (x) − λi | pi (x) < ε · pi (x) = ε. Finally, by Lemma 2.3, λi φ( pi ) ≤ 1, so that φ( f ) ≤ φ f − λi pi + λi φ( pi ) < 1 + ε · φ, and since ε was arbitrary, φ ≤ 1.

As an application of Theorem 2.4, we now prove an inequality due to von Neumann, which will be central to many later results. First we need a preliminary lemma. inθ be a strictly positive Lemma 2.5 (Fejer–Riesz). Let τ (eiθ ) = +N n=−N an e N pn z n function on the unit circle T. Then there is a polynomial p(z) = n=0 such that τ (eiθ ) = | p(eiθ )|2 . Proof. First note that since τ is real-valued, a−n = a¯ n and a0 is real. We may n+N , so that g is a polynomial of deassume a−N = 0. Set g(z) = +N n=−N an z iθ gree 2N with g(0) = 0. We have g(e ) = τ (eiθ ) · ei N θ = 0. Notice that the antisymmetry of the coefﬁcients of g implies g(1/¯z ) = z −2N g(z). This means that the 2N zeros of g may be written as z 1 , . . . , z N , 1/¯z 1 , . . . , 1/¯z N . We set q(z) = (z − z 1 ) · · · (z − z N ), h(z) = (z − 1/¯z 1 ) · · · (z − 1/¯z N ), and have that g(z) = a N q(z)h(z), with h(z) =

(−1) N z¯ N q(1/¯z ) . z1 · · · z N

Chapter 2. Positive Maps

13

Thus, τ (eiθ ) = e−i N θ g(eiθ ) = |g(eiθ )| = |a N | · |q(eiθ )| · |h(eiθ )| aN · |q(eiθ )|2 , = z ···z 1

N

N |1/2 q(z). so that τ (eiθ ) = | p(eiθ )|2 , where p(z) = | z1a···z N

N Writing p(z) = α0 + · · · + α N z N , we see that τ (eiθ ) = ,k=0 α α¯ k ei( −k)θ , so that the coefﬁcients of every strictly positive trigonometric polynomial have this special form. The above results can be shown to hold for positive trigonometric polynomials as well as strictly positive trigonometric polynomials. To do this, one must carefully examine the roots of the polynomial g that lie on the circle. Theorem 2.6. Let T be an operator on a Hilbert space H with T ≤ 1, and let S ⊆ C(T) be the operator system deﬁned by S = { p(eiθ ) + q(eiθ ): p, q are polynomials}. ¯ = p(T ) + q(T )∗ is positive. Then the map φ: S → B(H) deﬁned by φ( p + q) Proof. First, note that it is enough to prove that φ(τ ) is positive for every strictly positive τ . Indeed, if τ is only positive, then τ + 1 is strictly positive for every

> 0, and hence we have φ(τ ) + I = φ(τ + 1) ≥ 0 for every > 0, and it follows that φ(τ ) ≥ 0. Let τ (eiθ ) be strictly positive in S, so that τ (eiθ ) = +n ¯ k ei( −k)θ . We must prove that ,k=0 α α φ(τ ) =

+n

α α¯ k T ( − k)

,k=0

is a positive operator, where we deﬁne j T , j ≥ 0, T ( j) = T ∗− j , j < 0. To this end, ﬁx a vector x in our Hilbert space H and note that   ∗ ∗n    I .T .. . . T . α¯ 1 x T .. .. ..  α¯ 1 x    ..   ..  φ(τ )x, x =  .. . . . .  .  .  ,  . . . T∗  α¯ n x α¯ n x Tn ... T I

(∗)

14

Chapter 2. Positive Maps

where the matrix operator on the right is acting on the direct sum of n copies of the Hilbert space H(n) . Thus, if we can show that the matrix operator is positive, we shall be done. To this end, set   0 ... ... ... 0 ..  T . . . .  ..  . .  R = 0 .. .. . ,  .. . . . . . . . . . . . ..  0 ... 0 T 0 and note that R n+1 = 0, R ≤ 1. Using I to also denote the identity operator on H(n) , we see that the matrix operator in (∗) can be written as I + R + R 2 + · · · + R n + R ∗ + · · · + R ∗n = (I − R)−1 + (I − R ∗ )−1 − I. To see that this latter operator is positive, ﬁx h in H(n) , and let h = (I − R)y for y in H(n) . One obtains ((I − R)−1 + (I − R ∗ )−1 − I )h, h = y, (I − R)y + (I − R)y, y − (I − R)y, (I − R)y = y2 − Ry2 ≥ 0,

since R is a contraction. We can now give our second proof of von Neumann’s inequality.

Corollary 2.7 (von Neumann’s Inequality). Let T be an operator on a Hilbert space with T ≤ 1. Then for any polynomial p, p(T ) ≤ p, where p = sup | p(eiθ )|. θ

Proof. Note that the operator system S deﬁned in Theorem 2.6 is actually a ∗-algebra and separates points. Hence, by the Stone–Weierstrass theorem, S is dense in C(T). By Proposition 2.1, the map φ is bounded and hence extends to C(T). By Exercise 2.2, the extension will also be positive. Hence, φ is contractive by Theorem 2.4, from which the result follows.

Chapter 2. Positive Maps

15

Excercises 2.15 and 2.16 present two other proofs of von Neumann’s inequality. We let A(D) denote the functions that are analytic on D and continuous on D− . By the maximum modulus principle the supremum of such a function over D− is attained on T. Thus, we may regard A(D) as a closed subalgebra of C(T). Since the polynomials are dense in A(D), the above inequality guarantees that the homomorphism p → p(T ) extends to a homomorphism of A(D), and we denote the image of an element f simply by f (T ), so that one has f (T ) ≤ f for all f in A(D). Another consequence of Theorem 2.6 that we shall frequently use is that if a is an element of some unital C ∗ -algebra A, a ≤ 1, then there is a unital, positive map φ: C(T) → A with φ( p) = p(a). This observation is used in the following two results. Corollary 2.8. Let B, C be C ∗ -algebras with unit, let A be a subalgebra of B, 1 ∈ A, and let S = A + A∗ . If φ: S → C is positive, then φ(a) ≤ φ(1)·a for all a in A. Proof. Let a be in A, a ≤ 1. By Proposition 2.1 and Exercise 2.2, we may extend φ to a positive map on the closure S − of S. As remarked above, there is a positive map ψ: C(T) → B with ψ( p) = p(a). Since A is an algebra, the range of ψ is actually contained in S − . Clearly, the composition of positive maps is positive, so by Theorem 2.4, φ(a) = φ ◦ ψ(eiθ ) ≤ φ ◦ ψ(1)·eiθ = φ(1).

If φ(1) = 1 in the above, then φ is a contraction on A. It is somewhat surprising that φ need not be a contraction on all of S. We shall see an example of this phenomenon in Chapter 5. Corollary 2.9 (Russo–Dye). Let A and B be C ∗ -algebras with unit, and let φ: A → B be a positive map. Then φ = φ(1). Proof. Apply Corollary 2.8.

So far we have concentrated on positive maps without indicating how positive maps arise. We close this discussion with two such results. Unlike our previous results, we shall see (Exercise 2.9) that the hypothesis that our map is unital is crucial to the next three results.

16

Chapter 2. Positive Maps

Lemma 2.10. Let A be a C ∗ -algebra, S ⊆ A an operator system, and f : S → C a linear functional with f (1) = 1, f = 1. If a is a normal element of A and a ∈ S, then f (a) will lie in the closed convex hull of the spectrum of a. Proof. Suppose not. Note that the convex hull of a compact set is the intersection of all closed disks containing the set. Thus, there will exist a λ and r > 0 such that | f (a) − λ| > r , while the spectrum σ (a) of a satisﬁes σ (a) ⊆ {z: |z − λ| ≤ r }. But then σ (a − λ · 1) ⊆ {z: |z| ≤ r }, and since norm and spectral radius agree for normal elements, a − λ1 ≤ r , while | f (a − λ · 1)| > r . This contradic tion completes the proof. Since the convex hull of the spectrum of a positive operator is contained in the nonnegative reals, we see that Lemma 2.10 implies that such an f must be positive. Proposition 2.11. Let S be an operator system, B a unital C ∗ -algebra, and φ: S → B a unital contraction. Then φ is positive. Proof. Since B can be represented on a Hilbert space, we may, without loss of generality, assume that B = B(H) for some Hilbert space H. Fix x in H, x = 1. Setting f (a) = φ(a)x, x, we have that f (a) = 1, f ≤ φ. By Lemma 2.10, if a is positive, then f (a) is positive, and consequently, since x was arbitrary, φ(a) is positive. Proposition 2.12. Let A be a unital C ∗ -algebra, and let M be a subspace of A containing 1. If B is a unital C ∗ -algebra and φ: M → B is a unital contraction, ˜ M + M∗ → B given by then the map φ: ˜ + b∗ ) = φ(a) + φ(b)∗ , φ(a is well deﬁned and is the unique positive extension of φ to M + M∗ . ˜ then by the self-adjointness of Proof. If φ does extend to a positive map φ, ˜ positive maps (Exercise 2.1), φ necessarily satisﬁes the above equation. So we must prove that this formula yields a well-deﬁned, positive map. Note that to prove that φ˜ is well deﬁned, it is enough to prove that if a and a ∗ belong to M, then φ(a ∗ ) = φ(a)∗ . For this, set S1 = {a: a ∈ M and a ∗ ∈ M};

Spectral Sets

17

then S1 is an operator system, and φ is a unital, contractive map on S1 and hence positive by Proposition 2.11. Consequently by Exercise 2.1, φ is self-adjoint on S1 and so φ(a ∗ ) = φ(a)∗ for a in S1 . Thus, φ˜ is well deﬁned. To see that φ˜ is positive, it is sufﬁcient to assume that B = B(H), ﬁx x ˜ in H with x = 1, set ρ(a) ˜ = φ(a)x, x, and prove that ρ˜ is positive. Let ρ: M → C be deﬁned by ρ(a) = φ(a)x, x; then ρ = 1 and so, by the Hahn–Banach theorem, ρ extends to ρ1 : M + M∗ → C

with

ρ1 = 1.

But by Proposition 2.11, ρ1 is positive, and so ρ1 (a + b∗ ) = ρ(a) + ρ(b) = ρ(a ˜ + b∗ ). Hence, ρ˜ is positive. Note that the above result also shows that there is a unique, norm-preserving Hahn–Banach extension of ρ to M + M∗ . Example 2.13. A positive map need not have a positive extension unless the range is C (Exercise 2.10). Indeed, if the positive map of Example 2.2 had a positive extension to C(T), then by Corollary 2.9, this extension would be a contraction. If S is an operator system, contained in a C ∗ -algebra A, and φ is a linear functional on S with φ(1) = 1, then by Exercise 2.3 and Proposition 2.11, we see that φ is contractive if and only if φ is positive. These maps are called states on S.

Spectral Sets There is a natural way that positive maps on operator systems arise in operator theory, and that is in the study of spectral sets for operators. Let X be a compact set in the complex plane, and let R(X ) be the subalgebra of C(X ) consisting of quotients p/q of polynomials, where the zeros of q lie off X . Note that two quotients of polynomials can be equal as functions on X even though they may be distinct as elements of the formal algebra of quotients. Indeed, in the extreme case where X is a singleton, X = {λ}, then z − λ = 0 as a function on X . If T is in B(H), with the spectrum σ (T ) of T , contained in X , then for any quotient p/q as above we have an operator p(T )q(T )−1 . Thus, we can attempt to deﬁne a homomorphism ρ: R(X ) → B(H) by setting ρ( p/q) = p(T )q(T )−1 . In the extreme case that X = {λ}, this homomorphism will be well deﬁned if and only if T = λ · I.

18

Chapter 2. Positive Maps

If ρ is well deﬁned and ρ ≤ 1, then X is called a spectral set for T . When ρ is well deﬁned and only bounded with ρ ≤ K , then X is called a K-spectral set for T . We may also regard R(X ) as a subalgebra of C(∂ X ), the continuous functions on the boundary of X . By the maximum modulus theorem, this endows R(X ) with the same norm as if we regarded it as a subalgebra of C(X ). We let S = R(X ) + R(X ) regarded as a subset of C(∂ X ). Thus, f 1 + g1 = f 2 + g2 if and only if they are equal as functions on ∂ X , and are positive if and only if they are positive on ∂ X . The concept of spectral sets is partially motivated by von Neumann’s inequality, which can be interpreted as saying that an operator T is a contraction if and only if the closed unit disk is a spectral set for T . By Proposition 2.12, if X is a spectral set for T , then there is a well-deﬁned, positive map ρ: ˜ S → B(H) given by ρ( ˜ f + g¯ ) = f (T ) + g(T )∗ . Conversely, if the above map ρ˜ is well deﬁned and positive, then by Corollary 2.8, X is a spectral set for T .

Nonunital C ∗ -Algebras Let A be a nonunital C ∗ -algebra. We wish to make some comments on positive maps in this case. Let B be a C ∗ -algebra, and let φ: A → B be a positive map. We claim that φ is automatically bounded. To see this, note that it is enough to prove that φ is bounded on A+ . To this end, suppose that φ is not bounded; then there exists a sequence pn in A+ , pn ≤ 1, with φ( pn ) ≥ n 3 . Let p = n n −2 pn ; then we have that n −2 pn ≤ p and so n ≤ φ(n −2 pn ) ≤ φ( p) for all n, an obvious contradiction. Thus, φ is bounded. The second observation is that every nonunital C ∗ -algebra A embeds into a unital C ∗ -algebra, A1 [73]. Furthermore, positive maps from A to B extend to positive maps from A1 to B1 . To see this second statement, ﬁrst note that A is a closed, two-sided ideal in A1 , and so the map a + λ1 → λ, for a in A, is a ∗-homomorphism. Thus, if a + λ · 1 is positive, then λ ≥ 0. Now, let φ: A → B be positive; then since φ is bounded, φ( p) ≤ φ · 1 for all positive p with p ≤ 1. Deﬁne φ1 : A1 → B1 by φ1 (a + λ1) = φ(a) + λq where q = φ · 1. This map is positive, since if a + λ1 ≥ 0, then −λ−1 a ≤ 1, so that φ(−λ−1 a) ≤ q or 0 ≤ φ(a) + λq. Note that if we let A = B denote the compact operators on an inﬁnitedimensional Hilbert space, then there is no extension of the identity map to a positive map from A1 to B.

Toeplitz Matrices and Trigonometric Moments

19

Toeplitz Matrices and Trigonometric Moments We close this chapter with another application of the Fejer–Riesz lemma. A sequence of numbers {ak }+∞ k=−∞ is called a trigonometric moment sequence provided there exists a positive Borel measure µ on the circle (equivalently, on [0, 2π]) such that 2π eikθ dµ(θ) ak = 0

for all k in Z. Note that in this case a−k = a¯ k and a0 = µ([0, 2π ]) ≥ 0. Hence, µ is automatically bounded. We call µ a representing measure for {ak }+∞ k=−∞ . By the Riesz representation theorem, there is a one-to-one correspondence between positive, bounded Borel measures µ on the circle and positive linear maps φ: C(T) → C given by φ( f ) = f dµ. Thus, {ak }+∞ k=−∞ is a trigonometric moment sequence precisely when setting φ(z k ) = ak , k ∈ Z, extends to give a positive linear functional on C(T). We shall call an inﬁnite matrix (bi, j )i,∞j=0 formally positive if each of the ﬁnite matrices B N = (bi, j )i,N j=0 is positive. (Recall that in this book positive means positive semideﬁnite.) Note that we are making no requirements about boundedness of B. Theorem 2.14. The sequence {ak }+∞ k=−∞ is a trigonometric moment sequence if and only if the Toeplitz matrix (ai− j )i,∞j=0 is formally positive. Proof. If {ak }+∞ k=−∞ is a moment sequence with measure µ, and p0 , . . . , p N are scalars, then N ai− j p¯ j pi = | p(z)|2 dµ ≥ 0, i, j=0

where p(z) = p0 + · · · + p N z N . Hence (ai− j )i,∞j=0 is formally positive. Conversely, assume that (ai− j )i,∞j=0 is formally positive. Let S ⊆ C(T) be the span of z k , k ∈ Z, which is a dense operator system in C(T). Deﬁne φ: S → C by φ(z k ) = ak . We wish to prove that φ is a positive linear functional on S. As in the proof of Theorem 2.6, it will be enough to assume that τ ∈ S is strictly positive and prove that φ(τ ) ≥ 0. By the Fejer–Riesz lemma, τ (z) = i,N j=0 pi p¯ j z (i− j) for some scalars, p0 , . . . , p N . Hence φ(τ ) = i,N j=0 pi p¯ j ai− j ≥ 0 by the positivity of the matrix (ai− j )i,N j=0 . Now by Exercise 2.2, the positive map φ extends by continuity to a positive theorem we have a map on S − = C(T), and hence by the Riesz representation bounded, positive Borel measure with ak = φ(z k ) = z k dµ for all k in Z.

20

Chapter 2. Positive Maps

There is another characterization of trigonometric moment sequences that is often useful. If {ak }+∞ k=−∞ is a trigonometric moment sequence with representing measure µ, then it is easily seen that |ak | ≤ µ([0, 2π ]) = a0 and hence the power k series f (z) = a0 /2 + ∞ k=1 ak z converges to give an analytic function on the unit disk D. For |z| < 1 the Poisson kernel P(z, θ) =

1 1 + −1 iθ 1−e z 1 − e−iθ z

deﬁnes a positive function on the circle, and hence f (z) + f (z) = P(z, θ ) dµ(θ) ≥ 0. Thus, we have shown that if {ak }+∞ k=−∞ is a trigonometric moment sequence then f (z) is analytic on the disk and maps the disk into the right half plane. The converse of this result holds as well, and we sketch the proof. If f , as above, is analytic on D and has nonnegative real part, then gr (eiθ ) = f (r eiθ ) + f (r eiθ ), r < 1, is continuous on T and nonnegative. Hence the operator of multiplication by gr on L 2 (T) is a positive operator. But the matrix of this operator with respect to the standard orthonormal basis {eikθ }+∞ k=−∞ is (ai− j r |i|+| j| )i,+∞ j=−∞ . Letting r tend to 1 yields the formal positivity of the Toeplitz matrix (ai− j )i,+∞ j=0 .

Notes For a survey of the theory of positive maps, see [222]. The idea of using von Neumann’s inequality to prove the Russo–Dye result seems to have originated with Choi. The usual proof involves another important result of Russo and Dye, namely, that the extreme points of the unit ball of a unital C ∗ -algebra are the unitary elements [206] (see [173] for an elegant proof of this result). Lemma 2.10 is a minor adaptation of [8]. Von Neumann’s original proof of his inequality [152] ﬁrst showed that the inequality was met for the M¨obius transformations of the disk, and then reduced verifying the inequality for a general analytic function to this special case. A later proof by Heinz [114] is based on the classical Cauchy–Poisson formula. Most other presentations rely on Sz.-Nagy’s dilation theorem (Theorem 1.1). The Fejer–Riesz lemma can be found in [89]. Foias [99] has shown how particular von Neumann’s inequality is to the theory of Hilbert spaces. He proves that if a Banach space has the property

Exercises

21

that every contraction operator on the Banach space satisﬁes von Neumann’s inequality, then that space is necessarily a Hilbert space. A result of Ando [5] implies that von Neumann’s inequality holds for pairs of commuting contractions, that is, p(T1 , T2 ) ≤ sup{| p(z 1 , z 2 )|: |z i | ≤ 1, i = 1, 2}, where T1 and T2 are commuting contractions and p is an arbitrary polynomial in two variables. Thus, by Proposition 2.12, there is a positive map ¯ = p(T1 , T2 ) + q(T1 , T2 )∗ . φ( p + q) It would be interesting to know if a proof of this latter fact could be given along the lines of Theorem 2.6. Such a proof could perhaps shed some additional light on the rather paradoxical results of Crabbe and Davies [69] and of Varopoulos [236], that for three or more commuting contractions, the analogue of von Neumann’s inequality fails. We shall examine these ideas in detail in Chapter 5.

Exercises 2.1 Let S be an operator system, B be a C ∗ -algebra, and φ: S → B a positive map. Prove that φ is self-adjoint, i.e., that φ(x ∗ ) = φ(x)∗ . 2.2 Let S be an operator system, B be a C ∗ -algebra, and φ: S → B be a positive map. Prove that φ extends to a positive map on the norm closure of S. 2.3 Let S be an operator system, and let φ: S → C be positive. Prove that φ ≤ φ(1). [Hint: Given a, chosen λ, |λ| = 1, such that |φ(a)| = φ(λa).] 2.4 Let S be an operator system, and let φ: S → C(X ), where C(X ) denotes the continuous functions on a compact Hausdorff space X . Prove that if φ is positive, then φ ≤ φ(1). 2.5 (Schwarz inequality) Let A be a C ∗ -algebra, and let φ: A → C be a positive linear functional. Prove that |φ(x ∗ y)|2 ≤ φ(x ∗ x)φ(y ∗ y). 2.6 Let T be an operator on a Hilbert space H. The numerical radius of T is deﬁned by w(T ) = sup{|T x, x|: x ∈ H, x = 1}. Prove that if φ: S → B(H) is positive and φ(1) = 1, then w(φ(a)) ≤ a. 2.7 Let T be an operator on a Hilbert space. Prove that w(T ) ≤ 1 if and only if 2 + (λT ) + (λT )∗ ≥ 0 for all complex numbers λ with |λ| = 1. 2.8 Prove that w(T ) deﬁnes a norm on B(H), with w(T ) ≤ T ≤ 2w(T ). Show that both inequalities are sharp. 2.9 Let S be an operator system, B a C ∗ -algebra, and φ: S → B a linear map such that φ(1) is positive, φ(1) = φ. Give an example to show

22

Chapter 2. Positive Maps

that φ need not be positive. In a similar vein, show that if M is as in Proposition 2.12, and φ(1) is positive with φ(1) = φ, then φ˜ need not be well deﬁned. 2.10 (Krein) Let S be an operator system contained in the C ∗ -algebra A, and let φ: S → C be positive. Prove that φ can be extended to a positive map on A. 2.11 Prove that for the following element of Mn+1 ,  a0 0  a1 a0  . .  .. . . an . . .

 ... 0 . .. . ..   ≤ inf{a0 + a1 z + · · · + an z n + r (z)},  .. r (z) . 0  a1 a0

where r (z) is a polynomial whose lowest-degree term is strictly greater than n, and the latter norm is the supremum norm over the unit disk. 2.12 (Jorgensen) Show that C = { p 2 (t): p is a polynomial with real coef¯ p, q are polynomials}, ﬁcients} is dense in C([0, 1])+ . Let S = { p + q: ¯ = p(2) + q(2). Show that φ is positive on C and well and set φ( p + q) deﬁned on the dense subset S of C([0, 1]), but still does not extend to a positive map on C([0, 1]). Compare with Proposition 2.1 and Exercise 2.2. 2.13 Let X be a compact subset of C. Prove that if X is a ﬁnite set, then X is a spectral set for T if and only if T is a normal operator with σ (T ) contained in X . Prove that if X is a subset of R, then X is a spectral set for T if and only if T = T ∗ and σ (T ) is contained in X . Prove that if X is contained in the unit circle, then X is a spectral set for T if and only if T is a unitary with σ (T ) contained in X . 2.14 (von Neumann) Let X be a compact subset of C, with R(X ) dense in C(∂ X ). Prove that X is a spectral set for T if and only if T is normal and σ (T ) is contained in ∂ X . 2.15 In this exercise we give an alternative proof of von Neumann’s inequality. We assume that the reader has some familiarity with integration of operator-valued functions. Let T ∈ B(H) with T < 1, and let p and q denote arbitrary polynomials. (i) Let P(t; T ) = (1 − e−it T )−1 + (1 − eit T ∗ )−1 − 1, and show that P(t; T ) ≥ 0 for all t, 2π 1 it it (ii) Show that p(T ) + q(T )∗ = 2π 0 ( p(e ) + q(e ))P(t; T ) dt, (iii) Deduce von Neumann’s inequality. 2.16 (Wermer) In this exercise we give an alternative proof of von Neumann’s inequality that is only valid for matrices. We assume that the reader is familiar with the singular-value decomposition of a matrix. Let

Exercises

23

T ∈ Mn with T ≤ 1, and write T = U SV with U, V unitary and S = diag(s1 , . . . , sn ) a positive diagonal matrix, 0 ≤ si ≤ 1, i = 1, . . . , n. Deﬁne an analytic matrix-valued function T (z 1 , . . . , z n ) = U Z V where Z = diag(z 1 , . . . , z n ), |z i | ≤ 1, i = 1, . . . , n. Fix a polynomial p. (i) Let x, y ∈ Cn , and let f (z 1 , . . . , z n ) = p(T (z 1 , . . . , z n ))x, y. Deduce that f achieves its maximum modulus at a point where |z 1 | = · · · = |z n | = 1. Note that at such a point T (z 1 , . . . , z n ) is unitary. (ii) Deduce that sup|zi |≤1 p(T (z 1 , . . . , z n )) occurs at a point where T (z 1 , . . . , z n ) is unitary. (iii) Deduce that p(T ) ≤ sup p(W ) over W ∈ Mn , unitary. (iv) Show that for W unitary, p(W ) ≤ p∞ . (v) Deduce von Neumann’s inequality for T ∈ Mn . 2.17 Let T = ( a0 bc ) ∈ M2 , a = b. (i) Show that for any polynomial p, p(a) c( p(a) − p(b))/(a − b) p(T ) = . 0 p(b) (ii) Show that if |a| ≤ 1, |b| ≤ 1, f ∈ A(D), and pn is a sequence of polynomials with f − pn ∞ → 0, then f (a) c( f (a) − f (b))/(a − b) pn (T ) → 0 f (b) entrywise and hence in norm. ¯ Show that T ≤ (iii) Assume |a|, |b| < 1. Let ϕ(z) = (z − b)/(1 − bz). 1 if and only if |ϕ(a)|2 (|a − b|2 + |c|2 ) ≤ |a − b|2 . (iv) What can you say when |a|, |b| ≤ 1? 2.18 Let T = ( a0 ac ) ∈ M2 . (i) Show that for any polynomial p, p(a) cp (a) . p(T ) = 0 p(a) (ii) Show that if |a| < 1, f ∈ A(D), and pn is a sequence of polynomials with f − pn ∞ → 0, then f (a) c f (a) pn (T ) → 0 f (a) entrywise and hence in norm.

24

Chapter 2. Positive Maps

¯ Show that T ≤ 1 if and (iii) Let |a| < 1 and ϕ(z) = (z − a)/(1 − az). only if |c| ≤ 1 − |a|2 . (iv) What can you say when |a| = 1? (v) Let |a| < 1, set c = 1 − |a|2 , and use von Neumann’s inequality to deduce that if f ≤ 1, then | f (a)| ≤ (1 − | f (a)|2 )/(1 − |a|2 ). 2.19 Let a b c T = 0 a b ∈ M3 . 0 0 a Obtain necessary and sufﬁcient conditions for T ≤ 1. 2.20 (Korovkin) Let f ∈ C([0, 1]) and let gx (t) = (t − x)2 (i) Given > 0, show that there exists a constant c > 0 depending only on and f such that | f (t) − f (x)| ≤ + cgx (t)

for all

0 ≤ t, x ≤ 1.

[Hint: There is a δ > 0 such that | f (t) − f (x)| ≤ for all |t − x| < δ.] (ii) Let φ: C([0, 1]) → C([0, 1]) be a positive map with φ(1) = 1. Show that − − cφ(gx )(x) ≤ φ( f )(x) − f (x) ≤ + cφ(gx )(x), and deduce that φ( f ) − f ≤ + c supx |φ(gx )(x)|. (iii) Let φn : C([0, 1]) → C([0, 1]) be a sequence of positive maps. Prove that if φn ( f i ) − f i → 0 as n → ∞ for f i (t) = t i , i = 0, 1, and 2, then φn ( f ) − f → 0 as n → ∞ for all f ∈ C([0, 1]). 2.21 The Bernstein maps φn : C([0, 1]) → C([0, 1]) are deﬁned by n n k k φn ( f )(t) = t (1 − t)n−k . f n k k=0 (i) Verify that the Bernstein maps are positive maps with φn (1) = 1, 2 φn (t) = t, φn (t 2 ) = t 2 + t−t . [Hint: Use ( nk ) nk = ( n−1 ).] n k−1 (ii) Deduce that φn ( f ) − f ∞ → 0 for all f ∈ C([0, 1]) (iii) Deduce the Weierstrass theorem, i.e., prove that the polynomials are dense in C([0, 1]). 2.22 A sequence {an }∞ n=0 of complex numbers is called a Hausdorff moment sequence if there exists a positive (ﬁnite) Borel measure µ on [0,1] such 1 that an = 0 t n dµ(t) for all n. Set bn,m = nk=0 ( nk )(−1)k ak+m , for all n, m ≥ 0.

Exercises

25

(i) Assuming the existence of such a measure µ, show that bn,m = 1 m n t (t − 1) dµ(t), and deduce that necessarily bn,m ≥ 0 for all 0 n, m ≥ 0. (ii) Let P ⊆ C([0, 1]) denote the span of the polynomials, and deﬁne φ: P → C by setting φ(t n ) = an . Show that if bn,m ≥ 0 for all n, m ≥ 0 then φ is a positive map. (Caution: You do not know φ is continuous! You must use Exercises 2.20 and 2.21.) (iii) Prove that {an }∞ n=0 is a Hausdorff moment sequence if and only if bn,m ≥ 0 for all n, m ≥ 0.

Chapter 3 Completely Positive Maps

Let A be a C ∗ -algebra, and let M be a subspace. Then we shall call M an operator space. Clearly, Mn (M) can be regarded as a subspace of Mn (A), and we let Mn (M) have the norm structure that it inherits from the (unique) norm structure on the C ∗ -algebra Mn (A). We make no attempt at this time to deﬁne a norm structure on Mn (M) without reference to A. Thus, one thing that distinguishes M from an ordinary normed space is that it comes naturally equipped with norms on Mn (M) for all n ≥ 1. Later in this book we shall give a more axiomatic deﬁnition of operator spaces, at which time we shall begin to refer to subspaces of C ∗ -algebras as concrete operator spaces. For now we simply stress that by an operator space M we mean a concrete subspace of a C ∗ -algebra, together with this extra “baggage” of a well-deﬁned sequence of norms on Mn (M). Similarly, if S ⊆ A is an operator system, then we endow Mn (S) with the norm and order structure that it inherits as a subspace of Mn (A). As before, if B is a C ∗ -algebra and φ: S → B is a linear map, then we deﬁne φn : Mn (S) → Mn (B) by φn ((ai, j )) = (φ(ai, j )). We call φ n-positive if φn is positive, and we call φ completely positive if φ is n-positive for all n. We call φ completely bounded if supn φn is ﬁnite, and we set φcb = sup φn . n

Note that ·cb is a norm on the space of completely bounded maps. We use the terms completely isometric and completely contractive to indicate that each φn is isometric and that φcb ≤ 1, respectively. We note that if φ is n-positive, then φ is k-positive for k ≤ n. Also, φk ≤ φn for k ≤ n (Exercise 3.1). In this chapter we investigate some of the elementary properties of these classes of maps and prove some theorems about when positive maps are automatically completely positive. We begin by relating some of the results of the previous chapter to these concepts. 26

Chapter 3. Completely Positive Maps

27

Lemma 3.1. Let A be a C ∗ -algebra with unit, and let a and b belong to A. Then: (i) a ≤ 1 if and only if

1 a a∗ 1

is positive in M2 (A). (ii) [ a1∗ ab ] is positive in M2 (A) if and only if a ∗ a ≤ b. Proof. We represent A on a Hilbert space H via π : A → B(H) and set A = π(a). If A ≤ 1, then for any vectors x, y in H, I A x x , = x, x + Ay, x + x, Ay + y, y

y A∗ I y ≥ x2 − 2Ayx + y2 ≥ 0. Conversely, if A > 1, then there exist unit vectors x and y such that Ay, x < −1 and the above inner product will be negative. The proof of (ii) is similar and is left as an exercise [Exercise 3.2(ii)]. Proposition 3.2. Let S be an operator system, B a C ∗ -algebra with unit, and φ: S → B a unital, 2-positive map. Then φ is contractive. Proof. Let a ∈ S, a ≤ 1. Then 1 φ(a) 1 a = φ2 ∗ a 1 φ(a)∗ 1 is positive and hence φ(a) ≤ 1.

Proposition 3.3 (Schwarz inequality for 2-positive maps). Let A, B be unital C ∗ -algebras, and let φ: A → B be a unital 2-positive map. Then φ(a)∗ φ(a) ≤ φ(a ∗ a) for all a in A. Proof. We have that [ 10 a0 ]∗ [ 10 a0 ] = [ a1∗ aa∗ a ] ≥ 0 and hence 1 φ(a) ≥ 0. φ(a)∗ φ(a ∗ a) By Lemma 3.1(ii), we have the result.

28

Chapter 3. Completely Positive Maps

Proposition 3.4. Let A and B be C ∗ -algebras with unit, let M be a subspace of A, 1 ∈ M, and let S = M + M∗ . If φ: M → B is unital and 2-contractive ˜ S → B given by φ(a ˜ + b∗ ) = φ(a) + φ(b)∗ is (i.e., φ2 ≤ 1), then the map φ: 2-positive and contractive. Proof. Since φ is contractive, φ˜ is well deﬁned by Proposition 2.12. Note that ˜ 2 = (φ˜ 2 ). Since φ2 is contractive, we M2 (S) = M2 (M) + M2 (M)∗ and that (φ) ˜ have again by Proposition 2.12, that φ 2 is positive and so φ˜ is contractive by Proposition 3.2. Proposition 3.5. Let A and B be C ∗ -algebras with unit, let M be a subspace of A, 1 ∈ M, and let S = M + M∗ . If φ: M → B is unital and completely ˜ S → B is completely positive and completely contractive. contractive, then φ: Proof. We have that φ˜ n is positive, since φn is unital and contractive, and φ˜ n is contractive, since φ˜ 2n = (φ˜ n )2 is positive. We’ve glossed over one point in the above proof. Namely, we’ve identiﬁed M2n (S) with M2 (Mn (S)). It’s obvious how to do this: one simply “erases” the additional brackets in an element of M2 (Mn (S)). One must check, however, that the norms one deﬁnes are the same in each instance. One has that M2 (Mn (S)) inherits its norm from M2 (Mn (A)), while M2n (S) inherits its norm from M2n (A). However, this “erasure” operation deﬁnes a ∗-isomorphism between M2 (Mn (A)) and M2n (A); thus the norms are indeed the same. Now that we’ve seen some of the advantages of considering maps in these two classes, let’s begin by describing some maps that belong to them. Let A and B be C ∗ -algebras. First, note that if π: A → B is a ∗-homomorphism, then π is completely positive and completely contractive, since each map, πn : Mn (A) → Mn (B), is a ∗-homomorphism, and ∗-homomorphisms are both positive and contractive. For a second class of maps, ﬁx x and y in A and deﬁne φ: A → A by φ(a) = xay. Note that if (ai, j ) is in Mn (A), then φn ((ai, j )) = (xai, j y)  x 0  . . .  0 =  .. . .  . . 0 ...

   ... 0   y 0 ... 0 . . a11 . . . a1n  . . . . .  .. . . .  . ..    .. ..   0  .   . . . .. .  . .. ..  . 0 . 0 an1 . . . ann 0 ... 0 y 0 x

≤ x · (ai, j ) · y.

Schur Products and Tensor Products

29

Thus, φ is completely bounded, and φcb ≤ x · y. A similar calculation shows that if x = y ∗ , then φ is completely positive. Combining these two examples, we obtain the prototypical example of a completely bounded map. Namely, let H1 and H2 be Hilbert spaces, let vi : H1 → H2 , i = 1, 2, be bounded operators, and let π: A → B(H2 ) be a ∗-homomorphism. Deﬁne a map φ: A → B(H1 ) via φ(a) = v2∗ π(a)v1 . Then φ is completely bounded with φcb ≤ v1 · v2 , and if v1 = v2 , then φ is completely positive. We shall prove in Chapter 8 that all completely bounded maps have this form. In each of the above examples, we see that the completely positive maps are all completely bounded. This is always the case. Proposition 3.6. Let S ⊆ A be an operator system, let B be a C ∗ -algebra, and let φ: S → B be completely positive. Then φ is completely bounded and φ(1) = φ = φcb . Proof. Clearly, we have that φ(1) ≤ φ ≤ φcb , so it is sufﬁcient to show φcb ≤ φ(1). To this end, let A = (ai, j ) be in Mn (S) with A ≤ 1, and let In be the unit of Mn (A), i.e., the diagonal matrix with 1’s on the diagonal. Since In A A ∗ In is positive, we have that φ2n

In A A ∗ In

=

φn (In ) φn (A) φn (A)∗ φn (In )

is positive. Thus, by Exercise 3.2(iii), φn (A) ≤ φn (In ) = φ(1), which completes the proof.

Schur Products and Tensor Products As an application of the above result, we study the Schur product of matrices. If A = (ai, j ), B = (bi, j ) are elements of Mn , then we deﬁne the Schur product by A ∗ B = (ai, j · bi, j ). For ﬁxed A, this gives rise to a linear map, S A : Mn → Mn , via S A (B) = A ∗ B.

30

Chapter 3. Completely Positive Maps

In order to study this map, we need to recall a few facts about tensor products of matrices. Let A be in Mn and B be in Mm , so that A and B can be thought of as linear transformations on Cn and Cm , respectively. Then A ⊗ B is the linear transformation on Cn ⊗ Cm Cnm , which is deﬁned by setting A ⊗ B(x ⊗ y) = Ax ⊗ By and extending linearly. Writing A ⊗ B = (A ⊗ I )(I ⊗ B), it is easy to see that A ⊗ B = A · B. Let {e1 , . . . , en } and { f 1 , . . . , f m } be the canonical orthonormal bases for Cn and Cm , respectively. If we order our basis for Cnm by e1 ⊗ f 1 , e1 ⊗ f 2 , . . . , e1 ⊗ f m , e2 ⊗ f 1 , . . . , en ⊗ f m , then the matrix for A ⊗ B with respect to this ordered basis is given in block form by   a11 B . . . a1n B ..   ..  . . . an1 B . . . ann B This last matrix is often referred to as the Kronecker product of A and B. On the other hand, if we order our basis for Cnm by e1 ⊗ f 1 , e2 ⊗ f 1 , . . . , en ⊗ f 1 , e1 ⊗ f 2 , e2 ⊗ f 2 , . . . , en ⊗ f m , then the matrix for A ⊗ B with respect to this ordered basis is given in block form by   b11 A . . . b1m A ..   ..  . . , bm1 A . . . bmm A which is the Kronecker product of B and A. Since the two block matrices given above represent the same linear transformation, they are unitarily equivalent. The unitary matrix that implements this unitary equivalence is simply the permutation matrix corresponding to the given reordering of the basis vectors. Note that one obtained the (i, j)th block of the second matrix, bi j A, by taking the (i, j) entry of the (k, )th block, ak, B, of the ﬁrst matrix. We shall encounter this rearrangement of matrix entries repeatedly in this book. We’ll refer to it as the canonical shufﬂe. Now let A and B be in Mn , and deﬁne an isometry V : Cn → Cn ⊗ Cn by V (ei ) = ei ⊗ ei . A simple calculation shows that V ∗ (A ⊗ B)V = A ∗ B. To see this, note that V ∗ (A ⊗ B)V e j , ei = A ⊗ B(e j ⊗ e j ), (ei ⊗ ei )

= Ae j , ei · Be j , ei = ai, j · bi, j = A ∗ Be j , ei .

Schur Products and Tensor Products

31

Thus, S A (B) = V ∗ (A ⊗ B)V ≤ A · B and so S A ≤ A. Similarly, if (Bi, j ) is in Mk (Mn ), then (S A )k ((Bi, j )) = (V ∗ (A ⊗ Bi, j )V )  ∗   V 0 ... 0 ..  B11 . . . .. ..  . . .   .. 0 =  .. . .  A ⊗ . ..  . . . 0 Bn1 . . . 0 ... 0 V∗

   V 0 ... 0 . B1n  .. . . ..  ..   0 . .   . .. .. .   .. . . 0 Bnn 0 ... 0 V

and so, (S A )k ≤ A also. Thus, S A cb ≤ A. This estimate is not very good in general. For example, the identity map is the Schur product against the matrix of all 1’s, and this matrix has norm n. If A is positive, then we shall prove below that S A is completely positive. Hence, for positive matrices we can obtain S A cb explicitly, by invoking Proposition 3.5. Namely, S A = S A (I ) = S A cb = max{ai,i : i = 1, . . . , n}. It is more difﬁcult to calculate S A cb when A is not positive. We return to this topic in Chapter 8. Clearly, if one decomposes A = (P1 − P2 ) + i(P3 − P4 ) with Pi positive, then S A = (S P1 − S P2 ) + i(S P3 − S P4 ). Thus, S A cb ≤ S P1 cb + S P2 cb + S P3 cb + S P4 cb , and each of the right hand terms is given by the maximum diagonal entry of the corresponding matrix. However, we shall see in Chapter 8 that this estimate can be far from S A cb . The following characterizes when a Schur product map is completely positive. Theorem 3.7. Let A = (ai j ) be in Mn . Then the following are equivalent: (i) A is positive, (ii) S A : Mn → Mn is positive, (iii) S A : Mn → Mn is completely positive. Proof. Clearly (iii) implies (ii). Note that the matrix J , all of whose entries are 1, is positive, and that S A (J ) = A. Hence, (ii) implies (i). It remains to prove that (i) implies (iii). First note that if A and B are positive then A ⊗ B is positive. To see this, note that A ⊗ B = (A1/2 ⊗ B 1/2 )2 . Now if B ∈ Mn is positive, then ∗ 1/2 A ⊗ B 1/2 V S A (B) = V ∗ (A ⊗ B)V = A1/2 ⊗ B 1/2 V is positive. Hence, (i) implies (ii). To see that (i) implies (iii), let B = (Bi j ) ∈

32

Chapter 3. Completely Positive Maps

Mk (Mn ) be positive, write B = (X i j )∗ (X i j ), and observe that (S A )k (B) = (V ∗ (A ⊗ Bi j )V ) = Y ∗ Y, where Y = ((A1/2 ⊗ X i j )V ).

There is an analogous theory of Schur product maps on B(2 ) where we regard bounded operators as inﬁnite matrices. If we demand that A be a bounded positive operator, then the arguments above can be used to show that S A is completely positive. However, the requirement that A be bounded is very restrictive. For example, the identity map on B(2 ) is the Schur product against the inﬁnite matrix of all 1’s, and this is not the matrix of a bounded operator. We shall leave these more delicate questions until we return to Schur product maps in Chapter 8. We now return to results about general completely positive maps. The next result shows that for linear functionals, the adverb “completely” introduces nothing new. Proposition 3.8. Let S be an operator space, and let f : S → C be a bounded linear functional. Then f cb = f . Furthermore, if S is an operator system and f is positive, then f is completely positive. Proof. Let (ai, j ) be in Mn (S), and let x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) be unit vectors in Cn . We have that | f n ((ai, j ))x, y | = f (ai, j )x j y¯ i = f ai, j x j y¯ i i, j i, j ≤ f· ai, j x j y¯ i . i, j Thus, we must show that this latter element has norm less than (ai, j ). To see this, note that the above sum is the (1,1) entry of the product    y¯ 1 · 1 . . . y¯ n · 1   x1 · 1 0 . . . 0 a11 . . . a1n   0 ... 0  . .. ..   ..   ..  .. ..  .. . , . .  .  .  . an1 . . . ann xn · 1 0 . . . 0 0 ... 0 and that the outer two factors each have norm equal to one, since x and y were chosen to be unit vectors. To prove that f is completely positive whenever f is positive reduces to showing that f n ((ai, j ))x, x = f ( i, j ai, j x j x¯ i ) is positive whenever (ai, j ) is

Schur Products and Tensor Products

33

positive. But using the above product with x = y, we see that the summation that f is being evaluated at is the (1,1) entry of a positive matrix and hence is positive. Let X be a compact Hausdorff space, and let C(X ) be the C ∗ -algebra of continuous functions on X . Note that every element F = ( f i, j ) of Mn (C(X )) can be thought of as a continuous matrix-valued function and that multiplication and the ∗-operation in Mn (C(X )) are just the pointwise multiplication and ∗operation of the matrix-valued functions. Thus, one way to make Mn (C(X )) into a C ∗ -algebra is to set F = sup{F(x): x ∈ X }, and by the uniqueness of C ∗ -norms, this is the only way. With these observations, the following is a direct consequence of Proposition 3.8. Theorem 3.9. Let S be an operator space, and let φ: S → C(X ) be a bounded linear map. Then φcb = φ. Furthermore, if S is an operator system and φ is positive, then φ is completely positive. Proof. Let x ∈ X , and deﬁne φ x : S → C by φ x (a) = φ(a)(x). By the above observations, φn = sup φnx : x ∈ X = sup{φ x : x ∈ X } = φ. Similarly, φn ((ai, j )) is positive if and only if φnx ((ai, j )) is positive for all x ∈ X , from which the second statement follows. Thus, when the range C ∗ -algebra is commutative, the concepts of bounded and completely bounded coincide, as do positive and completely positive. A commutative domain is also enough to ensure that positive maps are completely positive, as we shall prove shortly. However, we show in Chapter 14 that it is not enough to guarantee that bounded maps are completely bounded. Lemma 3.10. Let ( pi, j ) be a positive scalar matrix, and let q be a positive element of some C ∗ -algebra B. Then (q · pi, j ) is positive in Mn (B). Proof. Straightforward.

Theorem 3.11 (Stinespring). Let B be a C ∗ -algebra, and let φ: C(X ) → B be positive. Then φ is completely positive. Proof. Let P(x) be positive in Mn (C(X )). We must prove that φn (P) is positive. Given ε > 0 and arguing as in Theorem 2.4, we obtain a partition of unity {u (x)}

34

Chapter 3. Completely Positive Maps

and positive matrices P = ( pi, j ) such that u (x)P < ε. P(x) − But φn (u · P ) = φn ((u · pi, j )) = (φ(u ) · pi, j ), which is positive by Lemma 3.10. Thus, φn (P), to within εφn P, is a sum of positive elements. Since Mn (B)+ is a closed set, we have that φn (P) is positive. The original proof of Theorem 3.11 [221] is quite different. As an immediate application of Theorem 3.11 we have the following matrixvalued version of von Neumann’s inequality. Corollary 3.12. Let T be an operator on a Hilbert space H with T ≤ 1, and let ( pi, j ) be an n × n matrix of polynomials. Then ( pi, j (T )) B(H(n) ) ≤ sup{( pi j (z)) Mn : |z| = 1}. ¯ = p(T ) + q(T )∗ extends to a positive map Proof. The map given by φ( p + q) φ: C(T) → B(H). By Theorem 3.11 this map is completely positive, and so by Proposition 3.6, φcb = φ(1) = 1. Hence, ( pi j (T )) B(H(n) ) = φn (( pi j )) ≤ ( pi j ) Mn (C(T)) , and the result follows.

We’ve seen that for a commutative domain or range, positivity implies complete positivity. We shall see that if the domain or range isn’t too badly noncommutative, then slightly stronger hypotheses than positivity will be enough to imply complete positivity. For now we restrict the domain. Lemma 3.13. Let A be a C ∗ -algebra. Then every positive element of Mn (A) is a sum of n positive elements of the form (ai∗ a j ) for some {a1 , . . . , an } ⊆ A. Proof. We remark that if we let R be the element of Mn (A) whose kth row is a1 , . . . , an and whose other entries are 0, then R ∗ R = (ai∗ a j ), so such an element is positive. Now let P be positive, so P = B ∗ B, and write B = R1 + · · · + Rn , where Rk is the kth row of B and 0 elsewhere. We have that P = B ∗ B = R1∗ R1 + · · · + Rn∗ Rn , since Ri∗ R j = 0 when i = j.

Schur Products and Tensor Products

35

We note that by the above lemma, to verify that φ: A → B is n-positive it is sufﬁcient to check that (φ(ai∗ a j )) is positive for all {a1 , . . . , an } in A. Theorem 3.14 (Choi). Let B be a C ∗ -algebra, let φ: Mn → B, and let {E i, j }i,n j=1 denote the standard matrix units for Mn . The following are equivalent: (i) φ is completely positive. (ii) φ is n-positive. (iii) (φ(E i, j ))i,n j=1 is positive in Mn (B). Proof. Clearly, (i) implies (ii), and since (E i, j )i,n j=1 is positive, (ii) implies (iii). Thus, we shall prove that (iii) implies (i). For this it is sufﬁcient to assume that B = B(H). Fix k, and let x1 , . . . , xk belong to H and B1 , . . . , Bk belong to Mn . By the above lemma, it is sufﬁcient to prove that i, j φ(Bi∗ B j )x j , xi is positive. n br,s, Er,s so that Write B = r,s=1 Bi∗ B j =

n

b¯ r,s,i br,t, j E s,t .

r,s,t=1

Set yt,r =

k j=1

br,t, j x j ; then

n n ∗ ¯ br,s,i br,t, j x j , xi φ(Bi B j )x j , xi = φ(E s,t ) r =1 s,t=1

i, j

=

n

i, j

φ(E s,t )yt,r , ys,r .

r =1 s,t

But for each r , this last sum is positive, since (φ(E s,t ))ns,t=1 is positive. Thus, we’ve expressed our original sum as the sum of n positive quantities. This completes the proof. By combining the results of this chapter with the technique used in the proof of Theorem 2.6, some fairly deep operator theory results can be readily obtained. Recall that if T is in B(H), then we deﬁne the numerical radius of T by w(T ) = sup{| T x, x |: x ≤ 1, x ∈ H}. The following result is essentially due to Berger [15]. Theorem 3.15. Let T be in B(H), and let S ⊆ C(T) be the operator system ¯ p, q polynomials}. The following are equivalent: deﬁned by S = { p + q:

36

Chapter 3. Completely Positive Maps

(i) w(T ) ≤ 1. (ii) The map φ: S → B(H) deﬁned by ¯ = p(T ) + q(T )∗ + ( p(0) + q(0))I φ( p + q) is positive. Proof. We ﬁrst show that (i) implies (ii). Let Rn be the n × n operator matrix whose subdiagonal entries are T and whose remaining entries are 0. It is not difﬁcult to show that w(Rn ) ≤ w(T ) (Exercise 3.13). Mimicking the ﬁrst part of the proof of Theorem 2.6, we see that the above map φ is positive, provided that the operator matrices 

2

 T  ..  . Tn

 T ∗ . . . T ∗n ..  .. .. . . .   .. .. . . T∗  ... T 2

(∗)

are positive for all n. Note that as in Theorem 2.6, Rnn+1 = 0 and so the matrix (∗) can be written as (I − Rn )−1 + (I − Rn∗ )−1 . Fix a vector x = (I − Rn )y, and compute ((I − Rn )−1 + (I − Rn∗ )−1 )x, x = 2y2 − 2 Re Rn y, y . Thus, (∗) is positive if and only if w(Rn ) ≤ 1 (Exercise 3.13 (iv)). If w(T ) ≤ 1, then w(Rn ) ≤ 1 and so (∗) is positive, which implies that φ is positive. Conversely, if φ is positive, then, since S is dense in C(T), φ will be completely positive by Theorem 3.11. Note that the matrix 

1

 z  .. . zn

   . . . z¯ n 1 0 ... 0 . . .. ..  . ..  . ..  1  0 z   =  .. . .  .. .. ..   . z¯ . . . 0 . 0 . . . 0 zn ... z 1     1 0 ... 0 . 1 ... 1  .. . ..  ..  0 z¯ .  ×  ..  .   .. . . .. . . . 0 1 ... 1 0 . . . 0 z¯ n z¯

(∗∗)

is positive in Mn (C(T)), and so its image under φn will be positive. But the image of (∗∗) under φn is (∗). Thus, (∗) is positive for all n, and hence w(Rn ) ≤ 1.

Module Mappings and Multiplicative Domains 37 √ Let x ∈ H, x = 1, and let y = (x ⊕ · · · ⊕ x)/ n be a unit vector in the direct sum of n copies of H. We have that n−1 | T x, x |, n n for all n. from which it follows that w(T ) ≤ n−1 Thus, w(T ) ≤ 1, which completes the proof of the theorem. 1 ≥ | Rn y, y | =

Note that if w(T ) ≤ 1 and φ is as above, then since φ is positive, φ = φ(1) = 2. Thus, if p is a polynomial, then p(T ) = φ( p) − p(0)I ≤ 3 p. In particular, if w(T ) ≤ 1, then the functional calculus can be extended from polynomials to the disk algebra A(D). Corollary 3.16 (Berger–Kato–Stampﬂi). Let T be in B(H) with w(T ) ≤ 1, and let f be in A(D) with f (0) = 0. Then w( f (T )) ≤ f . Proof. It is sufﬁcient to assume that f is a polynomial and that f ≤ 1. Let ¯ = φ be the map of Theorem 3.14 for T . We must show that the map ψ( p + q) p( f (T )) + q( f (T ))∗ + ( p(0) + q(0))I is positive. But if p + q¯ is positive, then p ◦ f + q ◦ f is positive, and thus, ¯ = p( f (T )) + q( f (T ))∗ + ( p(0) + q(0))I = φ( p ◦ f + q ◦ f ) ψ( p + q)

is positive. Corollary 3.17 (Berger). Let T be in B(H). Then w(T n ) ≤ w(T )n .

Proof. We may assume w(T ) = 1, but then applying Corollary 3.16, with f (z) = z n , yields the result. Actually, the proof of Theorem 3.15 yields a matrix-valued generalization of the Berger–Kato–Stampﬂi theorem. If w(T ) ≤ 1 and F = ( f i j ) is in Mn (A(D)) with F(0) = 0, then w(F(T )) ≤ F. Module Mappings and Multiplicative Domains In later chapters an increasingly important role will be played by module actions and module mappings. Several key facts about module mappings are consequences of the Schwarz inequality, and so we present them here. Let A be a unital C ∗ -algebra, and assume that C is a subalgebra with 1C = 1A . Then we can regard A as a left C-module with module action c ◦ a = ca. Similarly, we can regard A as a right C-module or a C-bimodule. If B is another C ∗ -algebra containing C as such a subalgebra and φ: A → B is linear, then we

38

Chapter 3. Completely Positive Maps

call φ a left C-module map provided φ(ca) = cφ(a) for every a ∈ A and c ∈ C. We deﬁne right C-module maps and C-bimodule maps analogously. One of our main interests is in ﬁnding conditions that guarantee a map φ is a module map. Theorem 3.18. Let A and B be unital C ∗ -algebras, and let φ: A → B be completely positive with φ(1) = 1. We have the following: (i) {a ∈ A: φ(a)∗ φ(a) = φ(a ∗ a)} = {a ∈ A: φ(ba) = φ(b)φ(a) for all b ∈ A} is a subalgebra of A, and φ is a homomorphism when restricted to this set. (ii) {a ∈ A: φ(a)φ(a)∗ = φ(aa ∗ )} = {a ∈ A: φ(ab) = φ(a)φ(b) for all b ∈ A} is a subalgebra of A, and φ is a homomorphism when restricted to this set. (iii) {a ∈ A: φ(a)∗ φ(a) = φ(a ∗ a) and φ(a)φ(a)∗ = φ(aa ∗ )} = {a ∈ A: φ(ab) = φ(a)φ(b) and φ(ba) = φ(b)φ(a) for all b ∈ A} is a C ∗ subalgebra of A, and φ is a ∗-homomorphism when restricted to this set. Proof. We prove (i). The proofs of (ii) and (iii) are similar. First, if a belongs to the set on the right, then choosing b = a ∗ yields φ(a ∗ a) = φ(a ∗ )φ(a) = φ(a)∗ φ(a), since φ is self-adjoint. Thus, a belongs to the set on the left. So assume φ(a)∗ φ(a) = φ(a ∗ a), and apply the Schwarz inequality to the map ∗ (2) φ and the matrix [ a0 b0 ]. We obtain ∗ ∗ φ(a) φ(b∗ ) φ(a) φ(b∗ ) a a a ∗ b∗ ≤ φ (2) ba bb∗ 0 0 0 0 and so

φ(a ∗ a) − φ(a)∗ φ(a) φ(a ∗ b∗ ) − φ(a)∗ φ(b∗ ) ≥ 0. φ(b)φ(a) − φ(ba) φ(b)φ(b∗ ) − φ(bb∗ )

Since the (1,1) entry of this matrix is 0, it follows that the (2,1) entry must be 0 as well. Thus, φ(b)φ(a) = φ(ba), and we have that the two sets are equal. The remaining claims are trivial. Corollary 3.19. Let A, B, and C be unital C ∗ -algebras, and assume that C is a C ∗ -subalgebra of both A and B with 1C = 1A = 1B . If φ : A → B is completely positive and φ(c) = c for every c ∈ C, then φ is a C-bimodule map. We call the set given in Theorem 3.18(i) the right multiplicative domain of φ and denote it by R. Note that if we deﬁne a right R-module action on B via

Exercises

39

b ◦ a = bφ(a) for b ∈ B, a ∈ R, then φ is a right R-module map. Similarly, we call the set in Theorem 3.18(ii) the left multiplicative domain of φ, denoting it by L, and the set in Theorem 3.18(iii) the multiplicative domain of φ, denoting it by C. We have that L = R∗ and C = R ∩ R∗ and that φ is a C-bimodule map. It is perhaps worth remarking that the above proof only used 4-positivity of the map φ. Notes Lemma 3.1 is an observation used in the work of Choi and Effros [48]. Proposition 3.5 appears in Arveson [6], which is the source for many of the applications of complete positivity to operator theory. Theorem 3.11 is due to Stinespring [221], where the term “completely positive” is introduced and used. Stinespring’s proof was measure-theoretic. Theorem 3.14 can be found in Arveson [9] and Choi [41, 43]. Theorem 3.15, Corollary 3.16, and Corollary 3.17 can be found in Berger [15], Kato [130], and Berger and Stampﬂi [17]. For some related work see [16], [118], and [239]. These ideas were further generalized by the theory of C p operators in Sz.-Nagy and Foias [231] (see Exercises 4.16 and 8.10). For an elementary proof of Corollary 3.16, see [172]. Further results in the direction of Exercise 3.10 can be found in Tomiyama’s survey article [235]. Exercise 3.6(iii) appears in Choi [43] with a different proof. Exercise 3.9(v) is Walter’s 3 × 3 matrix trick [240], which yields a simple proof of 3.9(vi). Exercise 3.11 is an unpublished result of Smith [215]. Theorem 3.18 and Corollary 3.19 are due to Choi [42]. Exercises 3.1 Prove that φn ≤ φk for n ≤ k and that if φk is positive, then φn is positive. 3.2 Let P, Q, A be operators on some Hilbert space H with P and Q positive. (i) Show that [ AP∗ QA ] ≥ 0 if and only if | Ax, y |2 ≤ P y, y · Qx, x

for all x, y in H. (ii) Prove Lemma 3.1(ii). (iii) Show that if [ AP∗ QA ] ≥ 0, then for any x in H we have that 0 ≤ (P + A + A∗ + Q)x, x ≤ ( P x, x + Qx, x )2 and hence P + A A∗ + Q ≤ (P1/2 + Q1/2 )2 . (iv) Show that if [ AP∗ PA ] ≥ 0, then A∗ A ≤ P · P and in particular A ≤ P. 3.3 Prove a nonunital version of Proposition 3.2.

40

Chapter 3. Completely Positive Maps

3.4 (Modiﬁed Schwarz inequality for 2-positive maps) Let A and B be C ∗ algebras, and φ: A → B 2-positive. Prove that φ(a)∗ φ(a) ≤ φ(1)φ(a ∗ a) and that φ(a ∗ b)2 ≤ φ(a ∗ a) · φ(b∗ b). (Hint: Consider [ 10 a0 ] and [ a0 b0 ].) 3.5 Let A be a C ∗ -algebra with unit. Show that the maps Tr, σ : Mn (A) → A deﬁned by Tr((ai, j )) = i ai,i and σ ((ai, j )) = i, j ai, j are completely positive maps. Deduce that if (ai, j ) ≤ 1, then i, j ai, j ≤ n. 3.6 (Choi) Let A be a C ∗ -algebra, let λ be a complex number with |λ| = 1, let Uλ be the unitary element of Mn (A) that is diagonal with u i,i = λi , and let Diag: Mn (A) → Mn (A) be deﬁned by Diag((ai, j )) = (bi, j ), where bi, j = 0, for i = j and bi,i = ai,i . (i) Show that Uλ∗ (ai, j )Uλ = (λ j−i ai, j ). (ii) By considering the nontrivial nth roots of unity, show the map : Mn (A) → Mn (A) deﬁned by (A) = n Diag(A) − A is completely positive. (iii) Show that the map : Mn → Mn deﬁned by (A) = (n − 1) Diag(A) − A is not completely positive. 3.7 Let A and B be C ∗ -algebras with unit, and let φ1 , φ2 : A → B be bounded linear maps with φ1 ± φ2 completely positive. Prove that φ2 cb ≤ φ1 (1). [Hint: Use Lemma 3.1 applied to a and −a and Exercise 3.2(iv).] 3.8 Let A be a C ∗ -algebra with unit. Deﬁne T1 , T2 : Mn (A) → Mn (A) by T1 ((ai, j )) = (bi, j ) where bi,i = a, , bi, j = 0, i = j, and T2 ((ai, j )) = ± to be 1 in the (ci, j ) where ci, j = a j,i . Fix k and , k = , and deﬁne Uk, (k, ) entry and ±1 in the (, k) entry and 0 elsewhere. (i) Show that T1 (A) − T2 (A) =

1 −∗ − U AUk, . 2 k= k,

(ii) Show that T1 (A) + T2 (A) =

1 +∗ + U AUk, + Diag(A). 2 k= k,

(iii) Deduce that T1 ± T2 are completely positive and that T2 cb ≤ n. (iv) By considering A = C, show that T2 cb = n. 3.9 Let A be a C ∗ -algebra, and let Aop denote the set A with the same norm and ∗-operation, but with a multiplication deﬁned by a ◦ b = ba. (i) Prove that Aop is a C ∗ -algebra. op (ii) Prove that M2 and M2 are ∗-isomorphic via the transpose map. (iii) Show that the identity map from A to Aop is always positive.

Exercises

41 op

(iv) Prove that the identity map from M2 to M2 is not 2-positive. (v) (Walter) Let U, V, and X be elements of A with U, V unitary. Prove that   I U X U ∗ I V  ≥ 0 if and only if X = U V. X∗ V ∗ I (vi) Prove that the identity map from A to Aop is completely positive if and only if A is commutative. 3.10 (Tomiyama) Let A and B be unital C ∗ -algebras, and let (ai, j ) be in Mn (A). 1/2 2 (i) Prove that (ai, j ) ≤ (ai, j ) ≤ ≤ n(ai, j ), i, j ai, j and give examples to show that all of these inequalities are sharp. (ii) Let M be an operator space in A, and let φ: M → B be bounded. Prove that φn ≤ nφ. 3.11 (Smith) Let M be an operator space, and let φ: M → Mn be bounded. Prove that φcb ≤ nφ. [Hint: Write φ(a) =

n

φi, j (a) ⊗ E i, j ,

i, j=1

where φi, j : M → C.] 3.12 Let X be a compact subset of C, and let R(X ) be the quotients of polynomials with poles off X . We may regard R(X ) as a subalgebra of C(∂ X ) or of C(X ). Prove that with respect to these two embeddings, the identity map from R(X ) to R(X ) is completely isometric. Use this result to deduce that the real part of a function in R(X ) is positive on X if and only if it is positive on ∂ X . 3.13 Let Sn denote the cyclic forward shift on Cn . That is, Sn e j = e j+1 (mod n), where e0 , . . . , en−1 is the canonical basis for Cn . (i) Show that Sn is unitarily equivalent to a diagonal matrix whose entries are the nth roots of unity. (ii) Let T be in B(H). Show that w(T ) = w(T ⊗ Sn ). (iii) Let Rn be the n × n operator matrix whose subdiagonal entries are T and that is 0 elsewhere. Show that w(Rn ) ≤ w(T ⊗ Sn ). (Hint: Consider xλ = λx1 ⊕ · · · ⊕ λn xn with |λ| = 1.) (iv) Show that Re Rn y, y ≤ 1 for all y = 1 if and only if w(Rn ) ≤ 1. 3.14 In this exercise we outline an alternative proof of Theorem 3.14. Let T ∈ B(H) with w(T ) < 1, and let p and q be arbitrary polynomials. (i) Show directly that σ (T ) is contained in the open unit disk. (Hint: Recall Exercise 2.7.)

42

Chapter 3. Completely Positive Maps (ii) Let Q(t; T ) = (1 − e−it T )−1 + (1 − eit T ∗ )−1 , and show that Q(t; T ) ≥ 0 for all t. (iii) Show that 1 p(T ) + q(T )∗ + ( p(0) + q(0))I = 2π

2π it it 0 ( p(e ) + q(e ))Q(t; T ) dt.

(iv) Deduce Theorem 3.14. n 3.15 Let {An }+∞ n=−∞ be a sequence in B(H). Prove that φ(z ) = An extends linearly to a completely positive map φ: C(T) → B(H) if and only if the n × n Toeplitz matrix (Ai− j ) is positive for all n. A sequence {An }+∞ n=−∞ satisfying these conditions is called an operator-valued trigonometric moment sequence. 3.16 Prove that the equivalent conditions of Exercise 3.15 are also equivalent to k the operator-valued power series F(z) = A0 /2 + ∞ k=1 Ak z converging ∗ on D and satisfying F(z) + F(z) ≥ 0 for |z| < 1. n 3.17 Let {An }∞ n=0 be a sequence in B(H). Prove that φ(t ) = An extends linearly to a completely positive map φ: C([0, 1]) → B(H) if and only if Bn,m = n n k k=0 ( k )(−1) Am+k ≥ 0 for all n, m ≥ 0. We call such a sequence an operator-valued Hausdorff moment sequence. 3.18 Let A be a unital C ∗ -algebra, and let (bi j ) be in Mn (A). Prove that (bi j ) is positive if and only if for every n-tuple (a1 , . . . , an ) of elements of A, we have that i,n j=1 ai bi j a ∗j is a positive element of A.

Chapter 4 Dilation Theorems

We saw our ﬁrst example of a dilation theorem in Chapter 1. Sz.-Nagy’s dilation theorem (Theorem 1.1) showed that every contraction operator on a Hilbert space H was the compression of a unitary operator on a Hilbert space K that contains H. In this chapter we focus on dilation theorems that characterize various classes of maps into B(H) as compressions to H of “nicer” maps into B(K), where K is a Hilbert space containing H. One of the most general dilation theorems of this type is Stinespring’s theorem, which characterizes completely positive maps from C ∗ -algebras into B(H) in terms of ∗-homomorphisms into B(K) for some other Hilbert space K. Theorem 4.1 (Stinespring’s dilation theorem). Let A be a unital C ∗ -algebra, and let φ: A → B(H) be a completely positive map. Then there exists a Hilbert space K, a unital ∗-homomorphism π: A → B(K), and a bounded operator V : H → K with φ(1) = V 2 such that φ(a) = V ∗ π(a)V. Proof. Consider the algebraic tensor product A ⊗ H, and deﬁne a symmetric bilinear function , on this space by setting a ⊗ x, b ⊗ y = φ(b∗ a)x, yH and extending linearly, where , H is the inner product on H. The fact that φ is completely positive ensures that , is positive semideﬁnite, since     x1 x1 n n aj ⊗ xj, ai ⊗ xi = φn ((ai∗ a j ))  ...  ,  ...  ≥ 0, j=1 i=1 xn xn H(n) 43

44

Chapter 4. Dilation Theorems

where , H(n) denotes the inner product on the direct sum H(n) of n copies of H, given by    x  y1 1  .  .  ,  ..  = x1 , y1 H + · · · + xn , yn H . . . (n) xn yn H Positive semideﬁnite bilinear forms satisfy the Cauchy–Schwarz inequality, |u, v|2 ≤ u, u · v, v. Thus, we have that {u ∈ A ⊗ H | u, u = 0} = {u ∈ A ⊗ H | u, v = 0 for all v ∈ A ⊗ H} is a subspace, N , of A ⊗ H. The induced bilinear form on the quotient space A ⊗ H/N deﬁned by u + N , v + N = u, v will be an inner product. We let K denote the Hilbert space that is the completion of the inner product space A ⊗ H/N . If a ∈ A, deﬁne a linear map π (a): A ⊗ H → A ⊗ H by

(aai ) ⊗ xi . π (a) ai ⊗ xi = A matrix factorization shows that the following inequality in Mn (A)+ is satisﬁed: (ai∗ a ∗ aa j ) ≤ a ∗ a · (ai∗ a j ), and consequently,

π(a) a j ⊗ x j , π (a) ai ⊗ xi = φ(ai∗ a ∗ aa j )x j , xi H ≤ a ∗ a · φ(ai∗ a j )x j , xi H i, j

= a2 ·

aj ⊗ xj,

i, j

ai ⊗ xi .

Thus, π(a) leaves N invariant and consequently induces a quotient linear transformation on A ⊗ H/N , which we still denote by π (a). The above inequality also shows that π(a) is bounded with π (a) ≤ a. Thus, π (a) extends to a bounded linear operator on K, which we still denote by π (a). It is straightforward to verify that the map π : A → B(K) is a unital ∗-homomorphism.

Chapter 4. Dilation Theorems

45

Now deﬁne V : H → K via V (x) = 1 ⊗ x + N . Then V is bounded, since V x2 = 1 ⊗ x, 1 ⊗ x = φ(1)x, xH ≤ φ(1) · x2 . Indeed, it is clear that V 2 = sup{φ(1)x, xH : x ≤ 1} = φ(1). To complete the proof, we only need observe that V ∗ π (a)V x, yH = π(a)1 ⊗ x, 1 ⊗ yK = φ(a)x, yH for all x and y, and so V ∗ π(a)V = φ(a).

There are several remarks to be made. First, any map of the form φ(a) = V ∗ π (a)V is easily seen to be completely positive. Thus, Stinespring’s dilation theorem characterizes the completely positive maps from any C ∗ -algebra into the algebra of bounded operators on any Hilbert space. Second, note that if φ is unital, then V is an isometry. In this case we may identify H with the subspace V H of K. With this identiﬁcation, V ∗ becomes the projection of K onto H, PH . Thus, we see that φ(a) = PH π(a)|H . If T is in B(K), then the operator PH T |H in B(H) is called the compression of T to H. If we decompose K = H ⊕ H⊥ and, using this decomposition, regard T as a 2 × 2 operator matrix, then the compression of T to H is just the (1,1) entry of this operator matrix for T. Thus, when φ(1) = 1, Stinespring’s theorem shows that every completely positive map into B(H) is the compression to H of a *-homomorphism into a Hilbert space that contains H. The third point to be made is that Stinespring’s theorem is really the natural generalization of the Gelfand–Naimark–Segal (GNS) representation of states. Indeed, if H = C is one-dimensional so that B(C) = C, then an isometry V : C → K is determined by V 1 = x and we have φ(a) = φ(a)1 · 1 = V ∗ π(a)V 1 · 1 = π (a)V 1, V 1K = π (a)x, x. In fact, rereading the above proof with H = C and A ⊗ C = A, the reader will ﬁnd a proof of the GNS representation of states. Finally, we note that if H and A are separable, then the space K constructed above will be separable as well. Similarly, if H and A are ﬁnite-dimensional, then K is ﬁnite-dimensional. We now turn our attention to considering the uniqueness of the Stinespring representation. We shall call a triple (π, V, K) as obtained in Stinespring’s

46

Chapter 4. Dilation Theorems

theorem a Stinespring representation for φ. Given a Stinespring representation (π, V, K), let K1 be the closed linear span of π (A)V H. It is easily veriﬁed that K1 reduces π(A) so that the restriction of π to K1 deﬁnes a ∗-homomorphism, π1 : A → B(K1 ). Clearly, V H ⊆ K1 , so we have that φ(a) = V ∗ π1 (a)V , i.e., that (π1 , V, K1 ) is also a Stinespring representation. It enjoys one additional property, namely, that K1 is the closed linear span of π1 (A)V H. Whenever the space of the representation enjoys this additional property, we call the triple a minimal Stinespring representation. The following result summarizes the importance of this minimality condition. Proposition 4.2. Let A be a C ∗ -algebra, let φ: A → B(H) be completely positive, and let (πi , Vi , Ki ),

i = 1, 2,

be two minimal Stinespring representations for φ. Then there exists a unitary U : K1 → K2 satisfying U V1 = V2 and U π1 U ∗ = π2 . Proof. If U exists, then necessarily, U π1 (ai )V1 h i = π2 (ai )V2 h i . i

i

Thus, it will be sufﬁcient to verify that the above formula yields a well-deﬁned isometry from K1 to K2 , since by the minimality condition, U will have dense range and hence be onto. To this end, note that 2 π1 (ai )V1 h i = V1∗ π1 (ai∗ a j )V1 h j , h i i i, j 2 φ(ai∗ a j )h j , h i = π2 (ai )V2 h i , = i, j i so U is isometric and consequently well deﬁned, which is all that we needed to show. We now show how some other dilation theorems can be deduced from Stinespring’s result. The following is a slightly reﬁned version of Theorem 1.1. It is interesting to compare the proofs.

Chapter 4. Dilation Theorems

47

Theorem 4.3 (Sz.-Nagy’s dilation theorem). Let T ∈ B(H) with T ≤ 1. Then there exists a Hilbert space K containing H as a subspace and a unitary U on K with the property that K is the smallest closed reducing subspace for U containing H such that T n = PH U n |H

for all nonnegative integers

n.

Moreover, if (U , K ) is another pair with the above properties, then there is a unitary V : K → K such that V h = h for h ∈ H and V U V ∗ = U . ¯ = p(T ) + q(T )∗ , Proof. By Theorem 2.6 and Exercise 2.2, the map φ( p + q) where p and q are polynomials, extends to a positive map of C(T) into B(H). This map is completely positive by Theorem 3.11. Let (π, V, K) be a minimal Stinespring representation of φ, and recall that since φ(1) = 1, we may identify V H with H. Setting π(z) = U , where z is the coordinate function, we have that U is unitary and that T n = φ(z n ) = PH π(z n )|H = PH U n |H . The minimality condition on (π, V, K) is equivalent to requiring that the span of {U n H: n = 0, ±1, ±2, . . . } be dense in K, which is equivalent to the requirement that there be no closed reducing subspace for U containing H other than K itself. The ﬁnal statement of the theorem is a consequence of the uniqueness of a minimal Stinespring representation up to unitary equivalence, Proposition 4.2.

The techniques used to prove Theorem 4.3 can be used to prove a far more general result. Let X ⊆ C be a compact set, and let R(X ) be the algebra of rational functions on X . An operator T is said to have a normal ∂ X -dilation if there is a Hilbert space K containing H as a subspace and a normal operator N on K with σ (N ) ⊆ ∂ X such that r (T ) = PH r (N )|H for all r in R(X ). We shall call N a minimal normal ∂ X -dilation of T , provided that K is the smallest reducing subspace for N that contains H. Clearly, when T has a normal ∂ X dilation N , r (T ) ≤ r (N ) ≤ sup{|r (z)|: z ∈ ∂ X },

48

Chapter 4. Dilation Theorems

and so a necessary condition for T to have a normal ∂ X -dilation is that X is a spectral set for T . It is a long-standing problem to determine if this condition is also sufﬁcient [231]. That is, if X is a spectral set for T , then does it necessarily follow that T has a normal ∂ X -dilation? No compact subsets of C are known where this fails to be true. Yet the collection of sets X for which this statement is known to always hold is somewhat limited. It has been veriﬁed that this condition is sufﬁcient for X an annulus [1], but the answer to this question is still unknown even when X is a nice region with two holes and T is a ﬁnite matrix. On the other hand, if T is restricted to be a 2 × 2 matrix, then this statement is known to be true [63] for every set X . If S = R(X ) + R(X ) is dense in C(∂ X ), then R(X ) is called a Dirichlet algebra on ∂ X . For example, if C/ X has only ﬁnitely many components and the interior of X is simply connected, then R(X ) is a Dirichlet algebra on ∂ X . See [68] for this and further topological conditions on X that ensure that R(X ) is a Dirichlet algebra on ∂ X . Theorem 4.4 (Berger–Foias–Lebow). Let R(X ) be a Dirichlet algebra on ∂ X . If X is a spectral set for T , then T has a minimal normal ∂ X -dilation. Moreover, any two minimal normal ∂ X -dilations for T are unitarily equivalent, via a unitary which leaves H invariant. Proof. Let ρ: R(X ) → B(H) be the unital contraction deﬁned by ρ(r ) = r (T ), so that ρ: ˜ S → B(H) is positive, where S = R(X ) + R(X ). Since S is dense in C(∂ X ) and positive maps are bounded, ρ˜ extends to a positive map φ on C(∂ X ). But by Theorem 3.11, φ is completely positive. The remainder of the proof proceeds as in Theorem 4.3. When R(X ) is not a Dirichlet algebra, minimal normal ∂ X -dilations of an operator need not be unitarily equivalent [165]. To state the next dilation theorem, we need to introduce some notation. Let X be a compact Hausdorff space, and let B be the σ -algebra of Borel sets on X . A B(H)-valued measure on X is a map E: B → B(H) that is weakly countably additive, that is, if {Bi } is a countable collection of disjoint Borel sets with union B, then E(Bi )x, y E(B)x, y = i

for all x, y in H. The measure is bounded provided that sup{E(B): B ∈ B} < ∞,

Chapter 4. Dilation Theorems

49

and we let E denote this supremum. The measure is regular provided that for all x, y in H, the complex measure given by µx,y (B) = E(B)x, y

(∗)

is regular. Given a regular bounded B(H)-valued measure E, one obtains a bounded, linear map φ E : C(X ) → B(H) via

φ E ( f )x, y =

f dµx,y .

(∗∗)

Conversely, given a bounded, linear map φ: C(X ) → B(H), if one deﬁnes regular Borel measures {µx,y } for each x and y in H by the above formula (∗∗), then for each Borel set B, there exists a unique, bounded operator E(B), deﬁned by the formula (∗), and the map B → E(B) deﬁnes a bounded, regular B(H)-valued measure. Thus, we see that there is a one-to-one correspondence between the bounded, linear maps of C(X ) into B(H) and the regular bounded B(H)-valued measures. Such measures are called (i) spectral if E(M ∩ N ) = E(M) · E(N ), (ii) positive if E(M) ≥ 0, (iii) self-adjoint if E(M)∗ = E(M), for all Borel sets M and N . Note that if E is spectral and self-adjoint, then E(M) must be an orthogonal projection and hence E is also positive. The following proposition, whose proof we leave to the reader (Exercise 4.10), summarizes the relationships between the above properties of measures and the properties of their corresponding linear maps. Proposition 4.5. Let E be a regular bounded B(H)-valued measure, and let φ: C(X ) → B(H) be the corresponding linear map. Then: (i) (ii) (iii) (iv)

φ is a homomorphism if and only if E is spectral; φ is positive if and only if E is positive; φ is self-adjoint if and only if E is self-adjoint; φ is a ∗-homomorphism if and only if E is self-adjoint and spectral.

The correspondence between these measures and linear maps leads to a dilation result for these measures.

50

Chapter 4. Dilation Theorems

Theorem 4.6 (Naimark). Let E be a regular, positive, B(H)-valued measure on X . Then there exists a Hilbert space K, a bounded linear operator V : H → K, and a regular, self-adjoint, spectral, B(K)-valued measure F on X , such that E(B) = V ∗ F(B)V. Proof. Let φ: C(X ) → B(H) be the positive, linear map corresponding to E. Then φ is completely positive by Theorem 3.11, and so we may apply Stinespring’s theorem to obtain a ∗-homomorphism π : C(X ) → B(K) and a bounded, linear operator V : H → K such that φ( f ) = V ∗ π ( f )V for all f in C(X ). If we let F be the B(K)-valued measure corresponding to π , then it is easy to verify that F has the desired properties. As another application of Stinespring’s theorem, we shall give a characterization of the completely positive maps between two matrix algebras. In contrast, there is an entire panoply of classes of positive maps between matrix algebras and much that is not known about the relationships between these various classes (see, for example [46] and [244]). We begin by remarking that if π : Mn → B(K) is a ∗-homomorphism, then up to unitary equivalence, K decomposes as an orthogonal, direct sum of ndimensional subspaces such that π is the identity representation on each of the subspaces (Exercise 4.11). Now let φ: Mn → Mk = B(Ck ), and let (π, V, K) be a minimal Stinespring representation for φ. By the construction of the space K given in Theorem 4.1, we see that dim(K) ≤ dim(Mn ⊗ Ck ) = n 2 k. Thus, up to unitary equivalence, we may decompose K into the direct sum of fewer than nk subspaces, each of dimension n, such that π : Mn → B(K) is the ⊕ Cin , ≤ nk, identity representation on each one. So let us write K = i=1 n and let Pi denote the projection of K onto Ci . We have that for any A in Mn , Pi π(A)|Cin = A. Now, if we let Vi : Ck → Cin be deﬁned by Vi = Pi V , then φ(A) = V ∗ π(A)V =

Vi∗ π(A)V j =

i, j=1

Vi∗ AVi ,

i=1

after identifying each Cin with Cn . We summarize these calculations in the following: Proposition 4.7 (Choi). Let φ: Mn → Mk be completely positive. Then there exist fewer than nk linear maps, Vi : Ck → Cn , such that φ(A) = i Vi∗ AVi for all A in Mn .

Chapter 4. Dilation Theorems

51

There is another dilation theorem due to Naimark whose proof is closely related to the proof of Stinespring’s theorem. Let G be a group and let φ: G → B(H). We call φ completely positive deﬁnite if for every ﬁnite set of elements g1 , . . . , gn of G, the operator matrix (φ(gi−1 g j )) is positive. If G is a topological group, then we call a map φ: G → B(H) weakly continuous provided φ(gλ )x, y → φ(g)x, y for every pair of vectors x, y in H, and every net {gλ } in G converging to g. Similarly, φ is called strongly continuous provided φ(gλ )x − φ(g)x → 0, and ∗-strongly continuous provided φ(gλ )x − φ(g)x → 0 and φ(gλ )∗ x − φ(g)x → 0. Theorem 4.8 (Naimark). Let G be a topological group, and let φ: G → B(H) be weakly continuous and completely positive deﬁnite. Then there exists a Hilbert space K, a bounded operator V : H → K, and a ∗-strongly continuous unitary representation ρ: G → B(H) such that φ(g) = V ∗ ρ(g)V. Consequently, φ is automatically ∗-strongly continuous. Proof. Consider the vector space C0 (G, H) of ﬁnitely supported functions from G to H and deﬁne a bilinear function on this space by φ(g −1 g ) f 1 (g ), f 2 (g)H . f1, f2 = g,g

As in the proof of Stinespring’s theorem, we have that f, f ≥ 0 and that the set N = { f | f, f = 0} is a subspace of C0 (G, H). We let K be the completion of C0 (G, H)/N with respect to the induced inner product. For h in H, deﬁne V h by h if g = e, (V h)(g) = 0 if g = e, where e denotes the identity of G, and let ρ: G → B(K) be left translation, that is, (ρ(g) f )(g ) = f (g −1 g ). It is straightforward to check that V is bounded and linear, that ρ is a unitary representation, and that φ(g) = V ∗ ρ(g)V . Now we show that ρ is weakly continuous. Let {gλ } be a net in G that converges to g0 . Since ρ is a unitary representation, it will sufﬁce to show that ρ(gλ ) converges weakly to ρ(g0 ) on a dense subspace.

52

Chapter 4. Dilation Theorems

Let f 1 , f 2 be in C0 (G, H). Then ρ(gλ ) f 1 , f 2 = φ(g −1 g ) f 1 gλ−1 g , f 2 (g)H g,g

= φ(g −1 gλ g ) f 1 (g ), f 2 (g)H , g,g

which, since both sums involved are ﬁnite, converges to φ(g −1 g0 g ) f 1 (g ), f 2 (g)H = ρ(g0 ) f 1 , f 2 . g,g

Thus, we see that ρ(gλ ) converges weakly to ρ(g0 ). Now it is easily checked that if a net of unitaries converges in the weak operator topology to another unitary, then it converges ∗-strongly. Thus, ρ is ∗-strongly continuous. The ∗-strong continuity of φ now follows from the representation and the ∗-strong continuity of ρ. It is useful to consider the special case of the above theorem when G = Z, the group of integers. Setting φ(n) = An , we see that φ is completely positive deﬁnite if and only if (Ani −n j )i,k j=1 is positive for every choice of ﬁnitely many integers n 1 , . . . , n k with k arbitrary. Taking n i = i, we see that this implies the formal positivity of (Ai− j )i,∞j=0 = T . A little careful reﬂection shows that an arbitrary choice of n 1 , . . . , n k is simply restricting T to the entries represented by the set {n 1 , . . . , n k } and then permuting these entries. Thus, a completely positive deﬁnite function on Z is just an operator-valued trigonometric moment sequence (Exercise 3.15). Naimark’s dilation theorem now tells us that there exists a unitary U = ρ(1) and V such An = V ∗ U n V for all n. As with Stinespring’s representation, there is a minimality condition that guarantees the uniqueness of this representation up to unitary equivalence (Exercise 4.12). A map φ: G → B(H) will be called positive deﬁnite if for every choice of n elements g1 , . . . , gn of G, and scalars, α1 , . . . , αn , the operator α¯ i α j φ gi−1 g j i, j

is positive. We remark that this is equivalent to requiring that for every x in H, the map φx : G → C, given by φx (g) = φ(g)x, x, be completely positive deﬁnite. We caution the reader that our terminology is not standard. What we have chosen to call completely positive deﬁnite is usually called simply positive deﬁnite, and the concept that we have introduced above and called positive

Chapter 4. Dilation Theorems

53

deﬁnite is usually not introduced at all. Our rationale for this slight deviation in notation will be clear in the section “Group C ∗ -Algebras” at the end of this chapter, where we will show a correspondence between the two classes of maps on G that are deﬁned above and maps on a certain C ∗ -algebra associated with G, C ∗ (G). Not surprisingly, this correspondence will carry (completely) positive deﬁnite maps on G to (completely) positive maps on C ∗ (G). We begin by describing this correspondence in one case of particular interest. Let Zn be the Cartesian product of n copies of the integers, and let Tn be the Cartesian product of n copies of the circle. Let J = ( j1 , . . . , jn ) be in Zn , and let j j z j be the jth coordinate function on Tn . We set z J = z 11 · · · z nn . Note that there is a one-to-one correspondence between unitary representations ρ: Zn → B(H) and ∗-homomorphisms π: C(Tn ) → B(H), given by π (z j ) = ρ(e j ), where e j is the n-tuple that is 1 in the jth entry and 0 in the remaining entries. Proposition 4.9. Let φ: Zn → B(H) be (completely) positive deﬁnite. Then there is a uniquely deﬁned, (completely) positive map ψ: C(Tn ) → B(H), given by ψ(z J ) = φ(J ). Conversely, if the (completely) positive map ψ is given, then the above formula deﬁnes a (completely) positive deﬁnite function φ on Zn . Proof. First, we consider the case where φ is completely positive deﬁnite. Let (ρ, V, K) be the Naimark dilation of φ, so that φ(J ) = V ∗ ρ(J )V , and let π: C(Tn ) → B(H) be the ∗-homomorphism associated with ρ. If we set ψ( f ) = V ∗ π ( f )V , then we obtain a map ψ: C(Tn ) → B(H), which is completely positive. Moreover, ψ satisﬁes ψ(z J ) = V ∗ π(z J )V = V ∗ ρ(J )V = φ(J ). The proof of the converse in the completely positive case is identical. Now, suppose that φ is only positive deﬁnite. If we ﬁx x in H and set φx (J ) = φ(J )x, x, then φx is a completely positive deﬁnite function on Zn . Thus, by the above there is a positive map ψx : C(Tn ) → C with ψx (z J ) = φx (J ). For ﬁxed f in C(Tn ), the function x → ψx ( f ) as x varies over H is a bounded, quadratic function (see Exercise 4.18), and hence there exists a bounded operator ψ( f ) such that ψ( f )x, x = ψx ( f ). This deﬁnes a linear map ψ: C(Tn ) → B(H), which is easily seen to be positive. The converse in the positive case is similar. Corollary 4.10. For Zn , the sets of positive deﬁnite and completely positive deﬁnite operator-valued functions coincide. Proof. Clearly, every completely positive deﬁnite map is positive deﬁnite. Now let φ: Zn → B(H) be positive deﬁnite. Then ψ: C(Tn ) → B(H), given ψ(z J ) = φ(J ) is a positive linear map. By Theorem 3.11, ψ is completely positive and hence φ is completely positive deﬁnite.

54

Chapter 4. Dilation Theorems

Again it is useful to understand what this corollary says in the case of Z . Setting φ(n) = An , we have seen that φ is completely positive deﬁnite if and only if (Ai− j )i,∞j=0 is formally positive. But φ is positive deﬁnite if and only if φx is positive deﬁnite for all x, which is if and only if (Ai− j x, x)i,∞j=0 is formally positive for all x. Thus, for Toeplitz operator matrices, formal positivity of (Ai− j ) is equivalent to formal positivity of (Ai− j x, x) for every vector x. Corollary 4.10 is a generalization of this observation to multivariable Toeplitz operators. We now wish to discuss the analogue for more general groups. Group C ∗ -Algebras The above results are part of a more general duality. Let G be a locally compact group, and let dg be a (left) Haar measure on G. The space L 1 (G) of integrable functions, on G can be made into a ∗-algebra by deﬁning f 1 (g) f 2 (g −1 g ) dg, f 1 × f 2 (g ) = and a ∗-operation by f ∗ (g) = (g)−1 f (g −1 ), where (·) is the modular function. It is then possible to endow L 1 (G) with a norm such that the completion of L 1 (G) is a C ∗ -algebra, denoted by C ∗ (G) (see [73] or [173]). There is a one-to-one correspondence between weakly continuous, unitary representations ρ: G → B(H), and ∗-homomorphisms π : C ∗ (G) → B(H), given by π( f ) = f (g)ρ(g) dg when f is in L 1 (G). See [173] for a development of this theory. In a similar fashion, there are one-to-one correspondences between the weakly continuous, (completely) positive deﬁnite, operator-valued functions on G and the (completely) positive maps on C ∗ (G), given by ψ( f ) = f (g)φ(g) dg for f in L 1 (G). The proof that the above formula deﬁnes a one-to-one correspondence between these two classes of maps is similar to the proof of Proposition 4.9 and is left as an exercise (Exercise 4.16). Proposition 4.9 follows from the above correspondence and the fact [173] that C ∗ (Z n ) = C(Tn ).

Exercises

55

Paralleling Corollary 4.10, we see that when G is commutative the positive deﬁnite and completely positive deﬁnite functions correspond. Notes Naimark’s two dilation theorems ([150] and [151]) preceded Stinespring’s dilation theorem [221]. Stinespring [221] deﬁned completely positive maps on C ∗ -algebras, proved that positive maps on C(X ) were completely positive (Theorem 3.11), and then proved Theorem 4.1 as a generalization of Naimark’s dilation theorem for positive, operator-valued measures. Arveson [6] realized the important role that the theory of completely positive maps can play in the study of normal ∂ X -dilations and gave the proofs of Theorems 4.3 and 4.4 that are presented here. Other proofs of Sz.-Nagy’s dilation theorem have used the theory of positive deﬁnite functions on Z or the “geometric” technique that we presented in Chapter 1, where the unitary and the space that it acts on are explicitly constructed ([228] and [210]). Two beautiful results of the geometric dilation techniques are Ando’s dilation theorem for commuting pairs of contractions [5] and the Sz.-Nagy–Foias commutant lifting theorem [231]. We will present these topics in Chapter 5. The correspondence between bounded, regular, operator-valued measures on a compact Hausdorff space X and bounded, linear maps on C(X ) is discussed in Hadwin [109]. Proposition 4.7 was proved by Choi [43], who also developed the theory of multiplicative domains [42] (Exercise 4.2). See Bunce [37] for more on Korovkin-type theorems (Exercise 4.9). Exercise 4.16 is from Berger [15]. The material distinguishing positive deﬁnite and completely positive deﬁnite functions on groups seems to be new. Exercises 4.1 Use Stinespring’s representation theorem to prove that V 2 = φcb when φ is completely positive. Also, use the representation theorem to prove that φ(a)∗ φ(a) ≤ φ(1)2 φ(a ∗ a). 4.2 (Multiplicative domains) In this exercise, we present an alternative proof of Theorem 3.18. Let A be a C ∗ -algebra with unit, and let φ: A → B(H) be completely positive, φ(1) = 1, with minimal Stinespring representation (π, V, K). (i) Prove that φ(a)∗ φ(a) = φ(a ∗ a) if and only if V H is an invariant subspace for π(a).

56

4.3

4.4

4.5

4.6 4.7

4.8

4.9

Chapter 4. Dilation Theorems (ii) Use this to give another proof that {a ∈ A: φ(a)∗ φ(a) = φ(a ∗ a)} = {a ∈ A: φ(ba) = φ(b)φ(a) for all b ∈ A}. Recall that this set is the right multiplicative domain of φ. (iii) Similarly, show that φ(a)∗ φ(a) = φ(a ∗ a) and φ(a)φ(a)∗ = φ(aa ∗ ) if and only if V H is a reducing subspace for π(a). Deduce that the set of all such elements is a C ∗ -subalgebra of A. Recall that this subalgebra is the multiplicative domain of φ. (Bimodule Maps) Let A, B, and C be C ∗ -algebras with unit, and suppose that C is contained in both A and B, with 1C = 1A and 1C = 1B . A linear map φ: A → B is called a C-bimodule map if φ(c1 ac2 ) = c1 φ(a)c2 for all c1 , c2 in C. Let φ: A → B be completely positive. (i) If φ(1) = 1, then prove that φ is a C-bimodule map if and only if φ(c) = c for all c in C. (ii) Prove, in general, that φ is a C-bimodule map if and only if φ(c) = c · φ(1) for all c in C. Moreover, in this case φ(1) commutes with C. Let Dn be the C ∗ -subalgebra of diagonal matrices in Mn . Prove that a linear map φ: Mn → Mn is a Dn -bimodule map if and only if φ is the Schur product map ST for some matrix T . Let φ: A → A be a completely positive projection with φ(1) = 1. (i) Show that if φ(a) = a, then φ(ax) = φ(aφ(x)) for every x. (ii) Show that B = {a: φ(a) = a} is a C ∗ -algebra, in the product a ◦ b = φ(ab), but that in general, B is distinct from the multiplicative domain of φ. (iii) Show that φ is a (B, ◦)-bimodule map. Let φ: G → B(H) be completely positive deﬁnite. Prove that φ is weakly continuous if and only if φ is strongly continuous. Let φ: G → Mn be continuous. Prove that φ is completely positive deﬁnite if and only if there exists a Hilbert space H and continuous functions xi : G → H, i = 1, . . . , n, such that φ(g −1 g ) = (x j (g ), xi (g)). A semigroup G with an involution g → g ∗ satisfying (g1 g2 )∗ = g2∗ g1∗ , 1∗ = 1 is called a ∗-semigroup. A function φ: G → B(H) is called completely positive deﬁnite if (φ(gi∗ g j )) is positive for every set of ﬁnitely many elements g1 , . . . , gn of G, and bounded if (φ(gi∗ g ∗ gg j )) ≤ Mg (φ(gi∗ g j )) where Mg is a constant depending only on g. Show that every group is a ∗-semigroup if we set g ∗ = g −1 . Prove a version of Naimark’s dilation theorem for ∗-semigroups. Let A be a C ∗ -algebra with unit, and let φn : A → B(H) be a sequence of completely positive maps such that φn (1) → 1 in the weak operator topology.

Exercises

57

(i) Prove that {a: φn (a)∗ φn (a) − φn (a ∗ a) → 0 and φn (a)φn (a)∗ − φn (aa ∗ ) → 0 in the weak operator topology} is a C ∗ -subalgebra of A. (ii) (Korovkin) Prove that if φn : C([0, 1]) → C([0, 1]) is a sequence of positive maps such that φn (1), φn (t), and φn (t 2 ) converge in norm to 1, t, and t 2 , respectively, then φn ( f ) converges in norm to f , for all f . 4.10 Prove Proposition 4.5. 4.11 Let π : Mn → B(K) be a unital ∗-homomorphism. Prove that up to unitary equivalence, K = H1 ⊕ · · · ⊕ Hn , with Hi = H, i = 1, . . . , n, such that π(E i, j ) is the identity map from H j to Hi . Show that up to unitary equivalence, π is the direct sum of dim(H) copies of the identity map. 4.12 Give a minimality condition for the Naimark representation of completely positive deﬁnite functions on a group, and prove uniqueness of minimal representations up to unitary equivalence. 4.13 Let t → A(t), t ≥ 0, be a weakly continuous semigroup of contraction operators, A(0) = I . For t < 0, set A(t) = A(−t)∗ . Prove that this extended map is completely positive deﬁnite in R. What does Naimark’s dilation theorem imply? (Hint: Recall the proof of Theorem 2.6.) 4.14 (Trigonometric moments) Let {An }+∞ n=−∞ be a sequence of bounded linear operators on a Hilbert space H. Prove that {An }+∞ n=−∞ is an operator-valued trigonometric moment sequence if and only if there exists a Hilbert space K, a unitary operator U on K, and a bounded linear operator V : H → K such that An = V ∗ U n V for all n. 4.15 (Hausdorff moments) Let {An }+∞ n=0 be a sequence of bounded linear operators on a Hilbert space H. Prove that {An }+∞ n=0 is an operator-valued Hausdorff moment sequence if and only if there exists a Hilbert space K, a positive contraction P on K, and a bounded linear operator V : H → K such that An = V ∗ P n V for all n ≥ 0. 4.16 Verify the claims of the subsection on group C ∗ -algebras. 4.17 (Berger) Let T be an operator on a Hilbert space H. Prove that w(T ) ≤ 1 if and only if there exists a Hilbert space K containing H and a unitary operator U on K such that T n = 2PH U n |H for all n ≥ 1. 4.18 Let H be a Hilbert space, and let γ : H → C be a function. We call γ quadratic provided that γ (λx) = |λ|2 γ (x) and γ (x + y) + γ (x − y) = 2(γ (x) + γ (y)) for all λ in C and for every x and y in H. If, in addition, there exists a constant M such that |γ (x)| ≤ Mx2 for all x in H, then we call γ bounded. Prove that γ is a bounded, quadratic function on H if and only if there exists T in B(H) such that γ (x) = T x, x, and that T is unique.

Chapter 5 Commuting Contractions on Hilbert Space

In this chapter we study families of commuting contractions on Hilbert space. For pairs of commuting contractions there is an analogue of Sz.-Nagy’s unitary dilation theorem due to Ando, and consequently a two-variable analogue of von Neumann’s inequality. Unlike the case of Sz.-Nagy’s and von Neumann’s theorems, there is really only one proof known of Ando’s dilation theorem, and this proof is geometric. It is similar in spirit to the proof of Sz.-Nagy’s theorem given in Chapter 1. Given this fact, it is somewhat surprising that the analogues of von Neumann’s inequality fail for three or more commuting contractions. This difference between the cases with two and with three (or more) variables is still not very well understood and is the source of a number of open questions. We begin with a dilation theorem for sets of commuting isometries. Theorem 5.1. Let {V1 , . . . , Vn } be a set of commuting isometries on a Hilbert space H. Then there is a Hilbert space K containing H and a set of commuting unitaries {U1 , . . . , Un } on K such that V1m 1 · · · Vnm n = PH U1m 1 · · · Unm n H for all sets {m 1 , . . . , m n } of nonnegative integers. Proof. Let U1 on K1 be the minimal unitary dilation of V1 given by Theorem 4.3. Recall that the span of {U1n H: n ∈ Z} is dense in K1 . For i = 1 we claim that there is a well-deﬁned isometry Wi : K1 → K1 given by the formula Wi

+N

U1n h n

=

n=−N

+N n=−N

58

U1n Vi h n .

Chapter 5. Commuting Contractions on Hilbert Space

59

To see this note that 2 +N n U1n−m Vi h n , Vi h m + U1 Vi h n = Vi h n , U m−n Vi h m n=−N n≥m n<m n−m = V1 Vi h n , Vi h m + Vi h n , V1m−n Vi h n n≥m

=

Vi V1n−m h n , Vi h m

n≥m

=

V1n−m h n , h m

n<m

+

Vi h n , Vi V1m−n h n

n<m

+

n≥m

n<m

n≥m

n<m

h n , V1m−n h n

U1n−m h n , h m + h n , U1m−n h n = 2 +N = U1n h n . n=−N

This equality of the norms proves that Wi is well deﬁned and an isometry. Note that if Vi is unitary, then Wi is onto a dense subspace of K1 and hence is also unitary. It is easy to see that {U1 , W2 , . . . , Wn } commute and that V1m 1 · · · Vnm n = PH U1m 1 W2m 2 · · · Wnm n H . Now continue by next taking the unitary dilation of W2 on K2 and extending U1 , W3 , . . . , Wn to be isometries on K2 . Since U1 is unitary, its extension will also be a unitary on K2 . Thus, after n such dilations and extensions we shall obtain an n-tuple of unitaries on a space Kn with the desired properties. We arrive at the ﬁrst of several results that rely on the above construction. Corollary 5.2. Let {V1 , . . . , Vn } be a set of commuting isometries on a Hilbert space H, and let pi j , i, j = 1, . . . , m, be polynomials in n variables. Then

( pi, j (V1 , . . . , Vn )) B(H(m) ) ≤ sup{ ( pi, j (z 1 , . . . , z n )) Mm : |z k | ≤ 1, 1 ≤ k ≤ n}. Proof. For U1 , . . . , Un as in Theorem 5.1 we have pi, j (V1 , . . . , Vn ) = PH pi j (U1 , . . . , Un )|H

60

Chapter 5. Commuting Contractions on Hilbert Space

and hence

( pi, j (V1 , . . . , Vn )) ≤ ( pi, j (U1 , . . . , Un )) , and the proof is completed by using the ∗-isomorphism of C ∗ ({U1 , . . . , Un }) with C(X ) for some compact subset X of the n-torus Tn . Corollary 5.3. Let {V1 , . . . , Vn } be a set of commuting isometries on a Hilbert space H. Then (V j∗ Vi ) ≥ (Vi V j∗ ) ≥ 0. Proof. Let {U1 , . . . , Un } be as in Theorem 5.1, and decompose K = H ⊕ H⊥ . Since each Vi is an isometry, Ui H ⊆ H, and hence relative to this decomposition, Vi X i Ui = . 0 Yi Since U ∗j Ui = Ui U ∗j , computing the (1, 1) entry of their operator matrix yields V j∗ Vi = Vi V j∗ + X i X ∗j . Hence, (V j∗ Vi ) = (Vi V j∗ ) + (X i X ∗j ) = (V1∗ , . . . , Vn∗ )∗ · (V1∗ , . . . , Vn∗ ) + (X 1∗ , . . . , X n∗ )∗ · (X 1∗ , . . . , X n∗ ), and both inequalities follow.

These inequalities allow us to generalize Theorem 5.1 considerably. Let G be an abelian group. We shall write the group operation as addition. We call P ⊆ G a spanning cone provided: (i) 0 ∈ P, (ii) if g1 , g2 ∈ P, then g1 + g2 ∈ P, (iii) if g ∈ G, then there exists g1 , g2 ∈ P such that g = g1 − g2 . Some key examples of abelian groups with spanning cones are G = Zn with P = (Z+ )n and any subgroup G ⊆ (R, +) with P = G ∩ R+ . Note that conditions (i) and (ii) make P a semigroup. We call ρ: P → B(H) a semigroup homomorphism if ρ(0) = I and ρ(g1 + g2 ) = ρ(g1 )ρ(g2 ). Theorem 5.4. Let G be an abelian group with spanning cone P, and let ρ: P → B(H) be a semigroup homomorphism such that ρ(g) is an isometry for every g ∈ P. Then there exists a Hilbert space K containing H and a unitary representation π: G → B(K) such that ρ(g) = PH π (g)|H for every g ∈ P.

Chapter 5. Commuting Contractions on Hilbert Space

61

Proof. Let g1 − g2 = g3 − g4 with gi ∈ P, i = 1, 2, 3, 4. Then ρ(g2 )∗ ρ(g1 ) = ρ(g2 )∗ ρ(g3 )∗ ρ(g3 )ρ(g1 ) = ρ(g2 + g3 )∗ ρ(g1 + g3 ) = ρ(g1 + g4 )∗ ρ(g1 + g3 ) = ρ(g4 )∗ ρ(g1 )∗ ρ(g1 )ρ(g3 ) = ρ(g4 )∗ ρ(g3 ). This calculation shows that there is a well-deﬁned map φ: G → B(H) given by setting φ(g) = ρ(g2 )∗ ρ(g1 ) where g = g1 − g2 with g1 , g2 in P. We claim that φ is completely positive deﬁnite. To see this, choose g1 , . . . , gn in G, and write gi = pi − qi where p1 , . . . , pn , q1 , . . . , qn are in P. We have that (φ(−gi + g j )) = (φ(− pi + qi + p j − q j )) = (ρ( pi + g j )∗ ρ(qi + p j )) = (ρ( pi )∗ ρ(q j )∗ ρ(qi )ρ( p j )) = D ∗ (ρ(q j )∗ ρ(qi ))D, where D is the diagonal matrix whose ( j, j) entry is ρ( p j ). Setting ρ(qi ) = Vi and applying Corollary 5.3, it is now easily seen that (φ(−gi + g j )) ≥ 0. Hence φ is completely positive deﬁnite. By Naimark’s dilation theorem there exists a Hilbert space K, an operator V : H → K, and a unitary representation π: G → B(K) such that φ(g) = V ∗ π(g)V . Since I = φ(0) = V ∗ V , we have that V is an isometry, and the proof is completed by identifying H with V H.

We now return to arbitrary commuting contractions. Theorem 5.5 (Ando’s dilation theorem). Let T1 and T2 be commuting contractions on a Hilbert space H. Then there exists a Hilbert space K that contains H as a subspace, and commuting unitaries U1 , U2 on K, such that T1n T2m = PH U1n U2m H for all nonnegative integers n, m. Proof. By Theorem 5.1 it will be enough to ﬁnd a pair of commuting isometries V1 , V2 such that T1n T2m = PH V1n V2m H (∗) The concrete isometric dilations of Chapter 1, Vi : 2 (H) → 2 (H) via Vi ((h 1 , h 2 , . . . )) = (Ti h 1 , Di h 1 , h 2 , . . . ), Di = (I − Ti∗ Ti )1/2 ,

i = 1, 2,

62

Chapter 5. Commuting Contractions on Hilbert Space

satisfy (∗), but these isometries need not commute. Indeed, we have V1 V2 ((h 1 , h 2 , . . . )) = (T1 T2 h 1 , D1 T2 h 1 , D2 h 1 , h 2 , . . . ) while V2 V1 ((h 1 , h 2 , . . . )) = (T2 T1 h 1 , D2 T1 h 1 , D1 h 1 , h 2 , . . . ). Assume for the moment that there exists a unitary U : H ⊕ H → H ⊕ H such that U ((D1 T2 h, D2 h)) = (D2 T1 h, D1 h). Let W : 2 (H) → 2 (H) be the unitary deﬁned by W ((h 1 , h 2 , . . . )) = (h 1 , U ((h 2 , h 3 )), U ((h 4 , h 5 )), . . . ). Note that W V1 , V2 W −1 are still isometries satisfying (∗) and that W V1 V2 W −1 ((h 1 , h 2 , . . . )) = W (T1 T2 h 1 , D1 T2 h 1 , D2 h 1 , U −1 ((h 2 , h 3 )), U −1 ((h 4 , h 5 )), . . . ) = (T1 T2 h 1 , U ((D1 T2 h 1 , D2 h 1 )), h 2 , h 3 , . . . ) = V2 V1 ((h 1 , h 2 , . . . )) = (V2 W −1 )(W V1 )((h 1 , h 2 , . . . )), and so W V1 , V2 W −1 commute. To obtain the desired unitary U , we note that

D1 T2 h 2 + D2 h 2 = [T2∗ (I − T1∗ T1 )T2 + (I − T2∗ T2 )]h, h = [T1∗ (I − T2∗ T2 )T1 + (I − T1∗ T1 )]h, h = D2 T1 h 2 + D1 h 2 . Thus, setting U ((D1 T2 h, D2 h)) = (D2 T1 h, D1 h) deﬁnes an isometric map between two subspaces of H ⊕ H. When the codimensions of these two subspaces agree, then this isometry extends to the desired unitary on H ⊕ H. This is the case for example when the dimension of H is ﬁnite. When H is inﬁnite-dimensional, however, these two codimensions can be different, and a more complicated argument is needed. One begins by redeﬁning the basic isometries, via Vi ((h 1 , h 2 , . . . )) = (Ti h 1 , Di h 1 , 0, h 2 , . . . ). Comparing V1 V2 with V2 V1 , one now ﬁnds that a unitary U : H(4) → H(4) is needed satisfying U ((D1 T2 h, 0, D2 h, 0)) = (D2 T1 h, 0, D1 h, 0).

Chapter 5. Commuting Contractions on Hilbert Space

63

The extra 0’s guarantee that when H is inﬁnite-dimensional then both subspaces are of the same inﬁnite codimension and so the desired U exists. Deﬁning W ((h 1 , h 2 , . . . )) = (h 1 , U ((h 2 , h 3 , h 4 , h 5 )), U ((h 6 , h 7 , h 8 , h 9 )), . . . ), one ﬁnds that W V1 and V2 W −1 commute and satisfy (∗).

Corollary 5.6. Let T1 and T2 be commuting contractions on a Hilbert space H, and let pi, j , i, j = 1, . . . , m, be polynomials in two variables. Then

( pi, j (T1 , T2 )) B(H(m) ) ≤ sup{ ( pi, j (z 1 , z 2 )) Mm : |z 1 | ≤ 1, |z 2 | ≤ 1}.

Proof. The proof is similar to Corollary 5.2.

We shall refer to the above result as the two-variable von Neumann inequality. Surprisingly, Ando’s construction cannot be generalized to more than two commuting contractions. For if it could, then one could prove an n-variable von Neumann inequality, and the analogue of von Neumann’s inequality fails for three or more commuting contractions. There are several such counterexamples in the literature. The following appeared in [236] and is perhaps the simplest. Example 5.7 (Kaijser–Varopoulos). Consider the following operators on C5 : 

0 0 0 0 1 0 0 0   0 0 0 A1 = 0  0 0 0 0 √ √ √ 0 1/ 3 −1/ 3 −1/ 3  0 0 0 0 0 0 0 0  1 0 0 0 A2 =   0 0 0 0 √ √ √ 0 −1/ 3 1/ 3 −1/ 3

 0 0  0 ,  0 0  0 0  0 ,  0 0

64

Chapter 5. Commuting Contractions on Hilbert Space

and 

0 0 0 0 0 0 0 0   0 0 0 A3 = 0  0 0 0 1 √ √ √ 0 −1/ 3 −1/ 3 1/ 3

 0 0  0 .  0 0

It is easily checked that Ai ≤ 1 and that Ai A j = A j Ai , for 1 ≤ i, j ≤ 3. If one considers the polynomial p(z 1 , z 2 , z 3 ) = z 12 + z 22 + z 32 − 2z 1 z 2 − 2z 1 z 3 − 2z 2 z 3 , then p ∞ = sup{| p(z 1 , z 2 , z 3 )|: |z i | ≤ 1} = 5, as a However,  0 0 0 0  0 0 0 0   0 0 0 0 p(A1 , A2 , A3 ) =   0 0 0  0 √ 3 3 0 0 0

little calculus shows.  0 0  0 ,  0 0

√ and since 3 3 > 5, the analogue of von Neumann’s inequality fails. Of course, once it fails for n = 3, it also fails for all n ≥ 3. In spite of the many examples of commuting contractions for which von Neumann’s inequality fails, it is still unknown whether or not it could hold up to some constant. That is, it is not known if for each n there exists a ﬁnite constant K n such that for any commuting contractions T1 , . . . , Tn and any polynomial p in n variables one has

p(T1 , . . . , Tn ) ≤ K n p ∞ , where p ∞ denotes the supremum of p over |z i | ≤ 1, 1 ≤ i ≤ n. It is generally believed that there does not exist such a constant. Some lower bounds √ on the possible values of K n are known. For example, it is known that K n ≥ 11n [75]. In a similar vein, one can consider the problem for matrices of polynomials, i.e., seek constants Cn such that for any commuting contractions T1 , . . . , Tn , one has

( pi, j (T1 , . . . , Tn )) ≤ Cn ( pi, j ) ∞ ,

Chapter 5. Commuting Contractions on Hilbert Space

65

where ( pi, j ) ∞ denotes the supremum of the norm of the matrix over |z i | ≤ 1, 1 ≤ i ≤ n, and the sizes of the matrices are arbitrarily large. Clearly, Cn ≥ K n , and so if there is no ﬁnite constant K n , then it should be easier to prove that there is no ﬁnite constant Cn ; but this is also still unknown. These questions lead naturally to our ﬁrst abstractly deﬁned operator algebra. Let Pn denote the algebra of polynomials in n variables. Given a polynomial p, we set

p u = sup p(T1 , . . . , Tn ) , where the supremum is taken over the family of all n-tuples of commuting contractions on all Hilbert spaces. It is easy to see that p u is ﬁnite, since it is bounded by the sum of the absolute values of the coefﬁcients of p, and that this quantity deﬁnes a norm on Pn . For a ﬁxed polynomial p, there is always an n-tuple of contractions where this supremum is acheived. To see this, ﬁrst choose a sequence of n-tuples (T , . . . , Tn,k ) such that p u = supk p(T1,k , . . . , Tn,k ) . If we then let Ti = 1,k k ⊕ Ti,k , i = 1, . . . , n, be the n-tuple of commuting contractions on the direct sum of the corresponding Hilbert spaces, then p(T1 , . . . , Tn ) = p u . Now, if for each polynomial p we choose a commuting n-tuple of contractions where p u is attained and form the direct sum of all such n-tuples, then we have a single n-tuple of commuting contractions (T1 , . . . , Tn ) on a Hilbert space H such that p u = p(T1 , . . . , Tn ) for every polynomial. Consequently, the map π : Pn → B(H) deﬁned by π ( p) = p(T1 , . . . , Tn ) is an isometric homomorphism of (Pn , · u ) into B(H). In a similar fashion, if for ( pi, j ) in Mk (Pn ) we set

( pi, j ) u,k = sup ( pi, j (T1 , . . . , Tn )) , then we obtain a sequence of norms on the matrices over Pn . We call (Pn , · u,k ) the universal operator algebra for n-tuples of commuting contractions. To justify this terminology, note that, arguing as above, we can choose a single Hilbert space H and a single n-tuple of commuting contractions (T1 , . . . , Tn ) on that space such that ( pi, j )) u,k = ( pi, j (T1 , . . . , Tn )) for every k and every k × k matrix of polynomials. The corresponding map π, will now be a completely isometric homomorphism of the matrix-normed algebra Pn into B(H). Hence, (Pn , · u,k ) can be realized completely isometrically as a concrete algebra of operators. We now turn our attention to further applications of Ando’s theorem. We begin with some re-formulations due to Sz.-Nagy and Foias [231].

66

Chapter 5. Commuting Contractions on Hilbert Space

Theorem 5.8 (Commutant lifting theorem). Let T be a contraction on a Hilbert space, and let (U, K) be the minimal unitary dilation of T . If R commutes with T , then there exists an operator S commuting with U such that R = S and RT n = PH SU n |H for n ≥ 0. Proof. By scaling we may assume that R = 1. Hence, by Ando’s theorem there exists commuting unitaries U1 , U2 on a Hilbert space K1 containing H such that T n R m = PH U1n U2m |H . Let K be the closed linear span of vectors of the form U1n h, n ∈ Z, h ∈ H. Then K is a reducing subspace for U1 , and U = PK U1 |K is the minimal unitary dilation of T . Relative to K1 = K ⊕ K⊥ we have operator matrices U 0 S B U1 = , U2 = . 0 V C D Since U1 and U2 commute, it follows that U and S commute and that PK U2 U1n |K = SU n . Thus PH SU n |H = PH U2 U1n H = RT n . Finally, 1 ≤ R ≤ S ≤ U2 = 1 and so R = S .

The following equivalent re-formulation of Ando’s result is also useful. Given Ti ∈ B(Hi ), i = 1, 2, and A ∈ B(H1 , H2 ), we say that A intertwines T1 and T2 provided that AT1 = T2 A. Note that this is equivalent to ( A0 00 ) commuting with ( T01 T02 ). Corollary 5.9 (Intertwining dilation theorem). Let Ti , i = 1, 2, be contraction operators on Hilbert spaces Hi with minimal unitary dilations (Ui , Ki ). If A intertwines T1 and T2 , then there exists R intertwining U1 and U2 such that

A = R and AT n = T n A = PH RU n = PH U n R 1

2

2

1 H1

2

2

H1

for all n ≥ 0. Proof. First note that the minimal unitary dilation of T1 0 Tˆ = 0 T2 is

U 0 1 Uˆ = 0 U2

acting on K1 ⊕ K2 .

Chapter 5. Commuting Contractions on Hilbert Space

67

Applying commutant lifting to Aˆ = ( A0 00 ), we obtain Rˆ = ( RA DB ) satisfying Aˆ Tˆ n = PH1 ⊕H2 Rˆ Uˆ n |H1 ⊕H2 . Equating the (2,1) entries of these operators yields AT n = PH2 RU1n |H1 for n ≥ 0. The other relations follow from the intertwining condition. ˆ = A ˆ = A and so A = R . Finally, A ≤ R ≤ R We now apply these results to characterize the norms of some operatorvalued matrices. First, some formalities are in order. Let’s assume that H is a separable Hilbert space and that f : T → H is measurable and square integrable. 2π Deﬁning h n ∈ H via h n = (1/2πi) 0 e−inθ f (eiθ ) dθ yields a formal Fourier +∞ inθ series n=−∞ e h n with 1 2π

2π

f (eiθ ) 2 dθ =

0

+∞

h n 2 .

n=−∞

This leads to identiﬁcations L 2 (T; H) = L 2 (T) ⊗ H = 2Z (H). If An ∈ B(H) is a norm-bounded sequence of operators and for 0 < r < 1 we set Br (eiθ ) =

+∞

An r |n| einθ ,

n=−∞

then the series converges uniformly to deﬁne a continuous operator-valued function Br : T → B(H). For f ∈ L 2 (T; H) we have that θ → Br (eiθ ) f (eiθ ) ∈ L 2 (T; H) and that multiplication by Br , M Br , is a bounded operator on L 2 (T; H) with

M Br = sup Br (eiθ ) ≡ Br ∞ θ

where Br (eiθ ) ∈ B(H). With respect to the identiﬁcation L 2 (T; H) = 2Z (H), the operator M Br becomes the doubly inﬁnite operator-valued Toeplitz matrix (r |i− j| Ai− j )i,+∞ j=−∞ . It can be veriﬁed that

(Ai− j ) = sup r |i− j| Ai− j = sup M Br = sup Br ∞ . r <1

r <1

r <1

Thus, when (Ai− j ) is ﬁnite, we identify this operator-valued Toeplitz ma inθ and write trix with the operator of multiplication by B(eiθ ) = +∞ n=−∞ An e

(Ai− j ) = B ∞ , even though the convergence of this latter series can be a touchy business. Thus, one should always interpret B ∞ ≡ supr <1 Br ∞ .

68

Chapter 5. Commuting Contractions on Hilbert Space

We shall need the following operator-valued version of Nehari’s theorem in Chapter 10. Theorem 5.10 (Nehari–Page). Let H be a separable Hilbert space, and let An ∈ B(H), n ≥ 0, be a sequence of operators. Then the operator-valued Hankel matrix (Ai+ j )i,∞j=0 is bounded on 2 (H) if and only if there exists A ∈ B(H), n < 0, such that B ∞ ≡ supr <1 Br < +∞, where B(eiθ ) = n+∞ inθ . Moreover, in this case there exists a particular choice of n=−∞ An e An ∈ B(H), n < 0, such that

(Ai+ j ) = B ∞ ,

B(eiθ ) =

+∞

An einθ .

n=−∞

Proof. We only prove that if H = (Ai+ j ) is bounded on 2 (H), then there exists An ∈ B(H), n < 0, such that H = B ∞ . To this end let S: 2 (H) → 2 (H) via S((h 0 , h 1 , . . . )) = (0, h 0 , h 1 , . . . ) with adjoint S ∗ ((h 0 , h 1 , . . . )) = (h 1 , h 2 , . . . ). These operators can be identiﬁed with the operator matrices   0 ... ...    I 0 . . . 0 I 0 ... ...   . ..   0 . . . S ∗ = 0 0 I S = 0 I , . .. .. . . . . . .  .. ..  . . 0 . . . . . .. .. . . . . . We have that S ∗ H = (Ai+1+ j ) = H S. It is straightforward to verify that the minimal unitary dilations of S and S ∗ are the forward and backward shifts Sˆ and Sˆ ∗ on 2Z (H). By Corollary 5.9 there exists R = (Ri, j ): 2Z (H) → 2Z (H) ˆ This implies that Ri, j = Ri+ j with R = H such that Sˆ ∗ R = R S. for some sequence Rn ∈ B(H), n ∈ Z, with Rn = An for n ≥ 0. Now let W : 2Z (H) → 2Z (H) be the unitary deﬁned by W ((. . . , h −1 , h 0 , h +1 , . . . )) = (. . . , h +1 , h 0 , h −1 , . . . ). Then RW = (Ri− j ) and hence

+∞ Rn einθ .

H = R = RW = n=−∞

∞

When the sequence {An } is just numbers, then the above theorem has a clearer inθ interpretation. We have B(eiθ ) = +∞ , and B ∞ is just the norm n=−∞ an e of B in L ∞ (T). Thus, the theorem identiﬁes the sequences {an }n≥0 which can

Notes

69

be “completed” by choosing {an }n≤−1 to be the Fourier coefﬁcients of an L ∞ function as those for which the corresponding Hankel matrix is bounded. Set 2π 1 f (eiθ )e−inθ dθ, fˆ(n) = 2πi 0 and let G = { f ∈ L ∞ (T): fˆ(n) = 0 for all n ≥ 0}. Then the map Q: L ∞ (T) → ˙ B(2 ), Q( f ) = ( fˆ(i + j))i,+∞ j=0 has kernel G, and by the above theorem Q: L ∞ (T)/G → B(2 ) is an isometric linear isomorphism onto the space of Hankel matrices. Alternatively, if we look at the map P: L ∞ (T) → B(2 ),

P( f ) = ( fˆ(−1 − i − j)),

then ker P = { f ∈ L ∞ : fˆ(n) = 0 for all n ≤ −1} ≡ H ∞ (T), and by the above ˙ L ∞ (T)/H ∞ (T) → B(2 ) is an isometric isomorphism theorem we have that P: onto the space of Hankel matrices. This last isomorphism gives rise to an important duality. Recall that the dual of L 1 (T) is L ∞ (T). If we let H 1 (T) = { f ∈ L 1 (T): fˆ(n) = 0 for all n ≤ −1}, then the annihilator of H 1 (T) is { f ∈ L ∞ (T): fˆ(n) = 0 for all n ≥ 0} = eiθ · H ∞ (T). Consequently we have isometric isomorphisms between the dual of H 1 (T) and L ∞ (T)/eiθ H ∞ (T) ∼ = L ∞ (T)/H ∞ (T), which is in turn isometrically isomorphic to the space of Hankel matrices. Explicitly, this dual pairing is given by A, f =

∞

an fˆ(n)

n=0

for A = (ai+ j )i,∞j=0 a bounded Hankel and f ∈ H 1 (T). Notes The fact that von Neumann’s inequality holds for two commuting contractions but not three or more is still the source of many surprising results and intriguing questions. Many deep results about analytic functions come from this dichotomy. For example, Agler [3] uses Ando’s theorem to deduce an analogue of the classical Nevanlinna–Pick interpolation formula for analytic functions on the bidisk. Because of the failure of a von Neumann inequality for three or more commuting contractions, the analogous formula for the tridisk is known to be false, and the problem of ﬁnding the correct analogue of the Nevanlinna–Pick formula for polydisks in three or more variables remains open.

70

Chapter 5. Commuting Contractions on Hilbert Space

Many results are now known to be equivalent to Ando’s theorem, but a truly independent argument for proving any of these results that doesn’t ultimately rest on dilating a pair of commuting operators has still not been found. For example, we shall see in Chapter 7 that Ando’s theorem is equivalent to the two-variable version of von Neumann’s inequality holding for matrices of polynomials. That is, Corollary 5.6 implies Ando’s dilation theorem. In Chapter 18, we shall see that Ando’s theorem is equivalent to a certain factorization formula holding for matrices of polynomials in two variables. Cole and Wermer [66] prove that Ando’s theorem is equivalent to another theorem about factoring polynomials in two variables. Agler’s two-variable Nevanlinna–Pick result is also equivalent to Ando’s theorem. Yet direct proofs of any of these results that do not rely on Ando’s theorem, or one of its close variants, have not been found. Moreover, because of the failure of the von Neumann inequality for three or more variables, each of the above-cited results is known to fail for three or more variables. This operator-theoretic approach to results in classical function theory, as represented by Theorem 5.10, Exercise 5.3, and the results cited above, owes much to Sarason [209] and the book of Sz.-Nagy and Foias [231]. This continues to be an active area of research with many engineering applications, particularly to what has come to be known as H ∞ -control theory. For an introduction to this area we recommend the text of Foias and Frazho [100]. It is remarkable that we do not know whether or not a constant exists, for each n, such that von Neumann’s inequality for n commuting contractions is true up to that constant. This illustrates our lack of understanding of these important topics.

Exercises 5.1 Prove that if (V1 , K1 ) and (V2 , K2 ) are any two minimal isometric dilations of a contraction operator T on H, then there exists a unitary U : K1 → K2 such that U h = h for all h ∈ H and U V1 U ∗ = V2 . 5.2 Let H1 , H2 be Hilbert spaces, let An ∈ B(H1 , H2 ) be a sequence of operators, and let A = (Ai+ j ): 2 (H1 ) → 2 (H2 ) be the corresponding Hankel operator. Prove the analogue of the Nehari–Page theorem in this setting. 5.3 (Carath´eodory’s completion theorem) Let a0 , . . . , an be in C. Use commutant lifting to prove that  a0 0 .   a 1 . .  . . ..  .. an . . .

 ... 0 .  .. . ..   .. . 0  a1 a0 M

n ∞ j j = inf ajz + bjz , j=0 j=n+1 ∞

n+1

Exercises

71

where the inﬁmum is over all sequences {b j } such that the resulting power series is bounded on D and the ∞-norm is the supremum over D. Moreover, there exists a sequence {b j } where the inﬁmum is attained. Thus, a polynomial can be completed to a power series whose supremum over the disk is bounded by 1 by adding higher-order terms if and only if the norm of the corresponding Toeplitz matrix is at most 1. Deduce that the map 

fˆ(0) 0 ˆ .  f (1) . . ∞  f ∈ H (T) →  . .. .  .. ˆf (n) . . .

... .. . .. . fˆ(1)

 0 ..  .   ∈ Mn+1  0  fˆ(0)

yields an isometric isomorphism of H ∞ (T)/ei(n+1)θ H ∞ (T) into Mn+1 . Generalize to the case where A0 , . . . , An are operators on a (separable) Hilbert space. [Hint: The above ﬁnite Toeplitz matrix commutes with 

0 ... ... ...  .. . 1  . .. . Sn =  . . 0. .. .. . . . . 0 ... 0 1

 0 ..  . ..  .] . ..   . 0

5.4 Let {T1 , . . . , Tn } be contractions on a Hilbert space H (possibly noncommuting). Prove that there exists a Hilbert space K containing H and unitaries {U1 , . . . , Un } on K such that Tik1 1 . . . Tikmm = PH Uik11 . . . Uikmm H where m, k1 , . . . , km are arbitrary nonnegative integers, and 1 ≤ i ≤ n for = 1, . . . , m. 5.5 (Schaffer [209]) Let T be a contraction on a Hilbert space H, and let 2Z (H) = +∞ n=−∞ ⊕ H denote the Hilbert space formed as the direct sum of copies of H indexed by the integers Z. Deﬁne an operator matrix ∗ 1/2 , U−1,0 = (I − U = (Ui j )i,+∞ j=−∞ by setting U0,0 = T, U0,1 = (I − T T ) ∗ 1/2 ∗ T T ) , U−1,1 = −T , Un,n+1 = I for n ≥ 1 or n ≤ −2, and Ui, j = 0 for all other pairs (i, j). Prove that U deﬁnes a unitary operator on 2Z (H) and that if we identify H with the 0th copy of H in 2Z (H), then T n = PH U n |H for all nonnegative integers n. 5.6 Fix n ≥ 1, and let Pn denote the algebra of polynomials in n variables. For each p in Pn we set p u = sup{ p(T1 , . . . , Tn ) }, where for now

72

Chapter 5. Commuting Contractions on Hilbert Space

the supremum is only taken over all commuting n-tuples of contractions {T1 , . . . , Tn } on the Hilbert space 2 . (i) Prove that if {T1 , . . . , Tn } are commuting contractions on any Hilbert space, then p(T1 , . . . , Tn ) ≤ p u and consequently this is the same norm as deﬁned earlier. (ii) Prove the corresponding result for matrices of polynomials. (iii) Prove that the universal operator algebra for n-tuples of commuting contractions (Pn , · u,k ) can be represented completely isometrically as an algebra of operators on a separable Hilbert space. 5.7 Prove that · u is a norm on Pn and that the completion of Pn in this norm is a Banach algebra. We denote this Banach algebra by Au (Dn ), whereas we let A(Dn ) denote the Banach algebra of functions that are continuous on the closed polydisk Dn and analytic on Dn , equipped with the supremum − norm, · ∞ . In fact, A(Dn ) ⊆ C(Dn ) is the closure of Pn in the supremum norm. 5.8 (i) Prove that Au (Dn ) ⊆ A(Dn ) as sets. (ii) Prove that Au (Dn ) = A(Dn ) if and only if there exists a constant K n such that p u ≤ K n p ∞ for every p in Pn . 5.9 Let K ⊆ B(2 ) denote the ideal of compact operators on 2 . Let A(Dn ; K) denote the set of continuous functions from the closure of Dn into K that are analytic on Dn , and let P(Dn ; K) ⊆ A(Dn ; K) denote the functions that are ﬁnite sums of the form z J K J with K J in K. If we identify Mk as the subspace of operators in K that are supported on the ﬁrst k × k block, then we may regard Mk (A(Dn )) ⊆ Mk+1 (A(Dn )) ⊆ A(Dn ; K). (i) Prove that A(Dn ; K) equipped with the supremum norm and pointwise product is a Banach algebra. (ii) Prove that P(Dn ; K) is dense in A(Dn ; K). (iii) Prove that the union over k of Mk (A(Dn )) is dense in A(Dn ; K). (iv) Give a deﬁnition of Au (Dn ; K) and prove the analogues of (i), (ii), and (iii). (v) Prove that there exists a constant Cn such that for all matrices of polynomials, ( pi, j ) u ≤ Cn ( pi, j ) ∞ if and only if Au (Dn ; K) = A(Dn ; K).

Chapter 6 Completely Positive Maps into Mn

In this chapter we characterize completely positive maps into Mn . This characterization allows us to prove an extension theorem for completely positive maps and lends further insight into the properties of positive maps on operator systems. These results are all consequences of a duality between maps into Mn and linear functionals. Let M be an operator space, and let {e j }nj=1 be the canonical basis for Cn . If A is in Mn , then let A(i, j) denote the (i, j) entry of A, so that A(i, j) = Ae j , ei . If φ: M → Mn is a linear map, then we associate to φ a linear functional sφ on Mn (M) by the following formula: sφ ((ai, j )) =

1 φ(ai, j )(i, j) . n i, j 2

Alternatively, if we let x denote the vector in Cn = Cn ⊕ · · · ⊕ Cn given by x = e1 ⊕ · · · ⊕ en , then sφ ((ai, j )) =

1 φn ((ai, j ))x, x, n 2

where the inner product is taken in Cn . We leave it to the reader to verify that φ → sφ deﬁnes a linear map from L(M, Mn ), the vector space of linear maps from M into Mn , into the vector space L(Mn (M), C). If M contains the unit and φ(1) = 1, then sφ (1) = 1. Finally, if s: Mn (M) → C, then we deﬁne φs : M → Mn via (φs (a))(i, j) = n · s(a ⊗ E i, j ), where a ⊗ E i, j is the element of Mn (M) which has a for its (i, j) entry and is 0 elsewhere. We leave it to the reader to verify that the maps φ → sφ and s → φs are mutual inverses. 73

74

Chapter 6. Completely Positive Maps into Mn

Theorem 6.1. Let A be a C ∗ -algebra with unit, let S be an operator system in A, and let φ: S → Mn . The following are equivalent: (i) φ is completely positive, (ii) φ is n-positive, (iii) sφ is positive. Proof. Obviously, (i) implies (ii). Also, that (ii) implies (iii) is clear by the alternative deﬁnition of sφ . So assume that sφ is positive, and we shall prove that φ is completely positive. By Krein’s theorem (Exercise 2.10), we may extend sφ from Mn (S) to a positive, linear functional s on Mn (A). Since s extends sφ , the map ψ: A → Mn associated with s extends φ. Clearly, if we can prove that ψ is completely positive, then φ will be completely positive. To verify that ψ is m-positive by Lemma 3.13, it is sufﬁcient to consider a positive element of Mm (A) of the form (ai∗ a j ). Since ψm ((ai∗ a j )) acts on Cmn , to see that it is positive, it is sufﬁcient to take x = x1 ⊕ · · · ⊕ xm , where each x j = k λ j,k ek is in Cn , and calculate ψ(ai∗ a j )x j , xi (∗) ψm ((ai∗ a j ))x, x = i, j

=

λ j,k λ¯ i, ψ(ai∗ a j )ek , e

i, j,k,

=

λ j,k λ¯ i, s(ai∗ a j ⊗ E ,k ).

i, j,k,

Let Ai be the n × n matrix whose ﬁrst row is (λi,1 , . . . , λi,n ) and whose remaining rows are 0. We have that λ¯ i, λ j,k E ,k , Ai∗ A j = k,

and thus (∗) becomes i, j

s(ai∗ a j ⊗ Ai∗ A j ) = s

i

∗ ai ⊗ Ai

aj ⊗ Aj

,

j

which is positive, since s is positive. Thus, ψ is m-positive for all m.

Theorem 6.2. Let A be a C ∗ -algebra with unit, S an operator system contained in A, and φ: S → Mn completely positive. Then there exists a completely positive map ψ: A → Mn extending φ.

Chapter 6. Completely Positive Maps into Mn

75

Proof. Let sφ be the positive, linear functional on Mn (S) associated with φ, and extend it to a positive, linear functional s on Mn (A) by Krein’s theorem. By Theorem 6.1, the map ψ associated with s is completely positive. Finally, since s extends sφ , it is easy to see that ψ extends φ. We now restate these results for operator spaces with unit. Theorem 6.3. Let A be a C ∗ -algebra with unit, let M be a subspace of A containing 1, and let φ: M → Mn with φ(1) = 1. The following are equivalent: (i) φ is completely contractive, (ii) φ is n-contractive, (iii) sφ is contractive. Proof. Let S = M + M∗ . Clearly, (i) implies (ii) and (ii) implies (iii). Assuming (iii), since sφ is unital and contractive, we may extend it to a positive, unital map s˜ φ on Mn (M) + Mn (M)∗ = Mn (S). By Theorem 6.1, the linear functional s˜ φ is associated with a completely positive map on S. This map is ˜ where φ(x ˜ + y ∗ ) = φ(x) + φ(y)∗ . Hence, φ˜ is completely readily seen to be φ, positive, and so φ must be completely contractive. Theorem 6.4. Let A be a C ∗ -algebra with unit, let M be a subspace of A containing 1, and let φ: M → Mn be an n-contractive map with φ(1) = 1. Then φ extends to a completely positive map on A. Proof. In the proof of Theorem 6.3, we saw that φ˜ is completely positive and hence extends to a completely positive map on A by Theorem 6.2. There is one way in which the above correspondence between linear functionals on Mn (S) and linear maps of S into Mn is not well behaved. Suppose that s: Mn (S) → C is positive and unital, so that s = 1. Then s gives rise to a completely positive map φs : S → Mn , but φs is not necessarily unital. Indeed, since φs (1)(i, j) = ns(1 ⊗ E i, j ) = ns(E i, j ), we have that φs is unital if and only if 1/n, i = j, s(E i, j ) = 0, i= j.

76

Chapter 6. Completely Positive Maps into Mn

Because of this fact, φs is not necessarily a contraction. In fact, it is not hard to construct examples of a unital, positive s such that φs = n (Exercise 6.1). To obtain generalizations of the above results to the case where Mn is replaced by B(H) requires some topological preliminaries, which we postpone until the next chapter. We turn instead to some other applications of this correspondence between linear functionals and linear maps. Let S be an operator system and S + its cone of positive elements. We deﬁne pi ⊗ Q i : pi ∈ S + , Q i ∈ Mn+ , S + ⊗ Mn+ = i

where the sum is ﬁnite. Note that S + ⊗ Mn+ is a cone contained in the cone Mn (S)+ of positive elements of Mn (S). We have seen above that if φ: S → Mn , then φ is completely positive if and only if its associated linear functional sφ is positive on Mn (S)+ . The following points out the relevance of the set deﬁned above. Lemma 6.5. Let φ: S → Mn . Then φ is positive if and only if sφ : Mn (S) → C assumes positive values on S + ⊗ Mn+ . Proof. Let φ: S → Mn be positive, let p be in S + , and let Q be in Mn+ . In order to prove that sφ assumes positive values on S + ⊗ Mn+ , it will be sufﬁcient to prove that sφ ( p ⊗ Q) is positive. Furthermore, since by Lemma 3.13 Q can be written as a convex sum of matrices of the form (α¯ i α j ), it will sufﬁce to assume that Q = (α¯ i α j ). Thus, p ⊗ Q = (α¯ i α j p) and φ(α¯ i α j p)(i, j) n · sφ ( p ⊗ Q) = i, j

=

α¯ i α j φ( p)e j , ei = φ( p)x, x ≥ 0,

i, j

where x = α1 e1 + · · · + αn en . Conversely, assume that sφ is positive on S + ⊗ Mn+ , let p be in S + , and let x = α1 e1 + · · · + αn en be a vector in Cn . We have that φ( p)x, x = φ(α¯ i α j p)e j , ei = n · sφ ((α¯ i α j p)) ≥ 0, i, j

since (α¯ i α j p) is in S + ⊗ Mn+ . Thus, φ is positive.

Theorem 6.6. Let S be an operator system. Then the following are equivalent: (i) every positive map φ: S → Mn is completely positive,

Chapter 6. Completely Positive Maps into Mn

77

(ii) every unital, positive map φ: S → Mn is completely positive, (iii) S + ⊗ Mn+ is dense in Mn (S)+ . Proof. Clearly, (i) implies (ii). The proof that (ii) implies (i) is left as an exercise (Exercise 6.2). We prove the equivalence of (i) and (iii). If S + ⊗ Mn+ is dense in Mn (S)+ and φ: S → Mn is positive, then, by Lemma 6.5, sφ will be positive on Mn (S)+ , and so, by Theorem 6.1, φ is completely positive. Conversely, if S + ⊗ Mn+ is not dense in Mn (S)+ , then ﬁx p in Mn (S)+ , but not in the closure of S + ⊗ Mn+ . By the Krein–Milman theorem, there will exist a linear functional s on Mn (S) that is positive on S + ⊗ Mn+ , but such that s( p) < 0. The linear map φs : S → Mn , induced by s, is then positive by Lemma 6.5, but not completely positive. Corollary 6.7. Let S be an operator system. Then the following are equivalent: (i) for every C ∗ -algebra B, every positive φ: S → B is completely positive, (ii) for every n, every positive φ: S → Mn is completely positive, (iii) S + ⊗ Mn+ is dense in Mn (S)+ for all n. Proof. Clearly, (ii) and (iii) are equivalent and (i) implies (ii). To see that (ii) implies (i), it is sufﬁcient to consider B = B(H). Given (ai, j ) in Mn (S)+ , to check that φn ((ai, j )) is positive, it is enough to choose x1 , . . . , xn in H and check that φ(ai, j )x j , xi ≥ 0. i, j

Let F be the ﬁnite-dimensional subspace spanned by these n vectors, and let ψ: S → B(F ) be the compression of φ to F. Identifying B(F ) with Mk , where k = dim(F ), we have that ψ is completely positive by (ii) and hence ψ(ai, j )x j , xi = φ(ai, j )x j , xi , 0≤ i, j

as desired.

i, j

As well as being related to complete positivity, the above cone S + ⊗ Mn+ determines the norm behavior of positive maps. Let S be an operator system, and let x be in S. We say that S has a partition of unity for x, and say that x is partitionable with respect to S, provided that for every ε > 0, there exist positive elements p1 , . . . , pn in S with i pi ≤ 1, and scalars λ1 , . . . , λn with

78

Chapter 6. Completely Positive Maps into Mn

|λi | ≤ x , such that

λi pi < ε. x − i

We say that S has a partition of unity for a subset M of S provided that every element of M is partitionable with respect to S. Lemma 6.8. Let S be an operator system and let x be in S, with x ≤ 1. Then the following are equivalent: (i) x is partitionable with respect to S, (ii) every positive map φ with domain S satisﬁes φ(x) ≤ φ(1) , (iii) [ x1∗ x1 ] is in the closure of S + ⊗ M2+ . Proof. The proof that (i) implies (ii) is identical to the proof of Theorem 2.4. To see that (ii) implies (iii), assume that (iii) is not met. Let s: M2 (S) → C be a linear functional such that

1 x s < 0, x∗ 1 while s is positive on S + ⊗ M2+ , and let φ: S → M2 be the linear map associated with s. By Lemma 6.5, φ is positive. Since

φ(1) φ(x) e e 1 x 1 1 = 2s , < 0, x∗ 1 e2 φ(x)∗ φ(1) e2 we have that

φ2

1 x∗

x 1

is not positive. If we construct a unital, positive map φ from φ as prescribed by Exercise 6.2(i), then by Exercise 6.2(ii) we have that

1 x 1 φ (x) = ∗ φ2 1 φ (x) x∗ 1 is not positive. Thus, by Lemma 3.1, φ (x) > 1 = φ (1) , and so (ii) does not hold. Now suppose that (iii) is true, let ε > 0, and let pi in S + and Q i in M2+ , i = 1, . . . , n, be such that

1 x (∗) − pi ⊗ Q i < ε. ∗ x 1 i

Chapter 6. Completely Positive Maps into Mn We have that

ri Qi = λ¯ i

λi ti

79

,

with ri ≥ 0,

ti ≥ 0

for all i.

We set si = (ri + ti )/2. Note that if si = 0, then λi = 0. The equation (∗) implies the following inequalities; ri pi < 1 + ε, 1−ε< i

1−ε<

ti pi < 1 + ε,

i

λi pi < ε. x − i Thus, 1−ε<

si pi < 1 + ε.

i

Setting λi /si = 0 when si = 0, we have (λi /si )si pi < ε, x − i with |λi /si | ≤ 1. These last two equations are clearly enough to guarantee that x is partitionable. Theorem 6.9. Let S be an operator system, M a subset of S. Then every positive map on S has norm φ(1) when restricted to M if and only if S has a partition of unity for M. Proof. The proof is an immediate application of the above lemma.

Corollary 6.10. Let B be a C ∗ -algebra with unit, let A be a subalgebra of B containing the unit, and let S = A + A∗ . Then S has a partition of unity for A. Proof. By Corollary 2.8, every positive map on S satisﬁes φ(a) ≤ φ(1) ·

a for a in A. Thus, by Theorem 6.9, S has a partition of unity for A. We now study two examples to illustrate the use of the above ideas. The second arises from the theory of hypo-Dirichlet algebras, which we shall examine in more detail in Chapter 11. A uniform algebra A on a compact Hausdorff

80

Chapter 6. Completely Positive Maps into Mn

space X is called hypo-Dirichlet provided that the closure of A + A¯ is a ﬁnitecodimension subspace of C(X ). Our ﬁrst example is just a reexamination of an earlier one. Let S and φ be the operator system and positive map of Example 2.2. Since φ(1) = 1, while

φ(z) = 2 with z = 1, we see that z is not partitionable with respect to S. It is interesting to attempt to show this directly. The second example is somewhat longer. Example 6.11. Let 0 < R1 < R2 , and let A denote the annulus in C with inner radius R1 and outer radius R2 , i.e., A = {z ∈ C: R1 ≤ |z| ≤ R2 }. We let R(A), as before, denote the algebra of rational functions on A, and let S denote the closure of R(A) + R(A) in C(∂A). By a theorem of Walsh [239], S is a codimension 1 subspace of C(∂A) and hence R(A) is a hypo-Dirichlet algebra. We let j = {z: |z| = R j }, j = 1, 2, denote the two boundary circles of A, and deﬁne positive linear functionals s j , j = 1, 2, on C(∂A) by 2π 1 f (R j eit ) dt. sj( f ) = 2π 0 k If f (z) = +n k=−n ak z is a ﬁnite Laurent polynomial, then s1 ( f ) = s2 ( f ) = a0 . Since the ﬁnite Laurent polynomials are dense in R(A) and since s1 and s2 are self-adjoint functionals, we have that s1 ( f ) = s2 ( f ) for all f in S. Since S is of codimension 1, we have that S = { f ∈ C(∂ A): s1 ( f ) = s2 ( f )}. We claim that S has no partition of unity for S. Before proving this claim, let us explore some of the consequences of the claim. Set 0, z ∈ 1 , p(z) = 1, z ∈ 2 , / S, and so the span so that s1 ( p) = 0 = 1 = s2 ( p). From this, we see that p ∈ of S and p is all of C(∂ A). We leave it to the reader to verify that the claim that S has no partition of unity for S yields a positive, unital map φ: S → M2 with the following properties (Exercise 6.4):

Chapter 6. Completely Positive Maps into Mn (i) (ii) (iii) (iv)

81

φ is not contractive. φ is not completely positive. φ has no positive extension to C(∂A). If L1 = { f ∈ S: f ≤ p}, L2 = { f ∈ S: f ≥ p}, then there is no matrix P in M2 satisfying φ( f 1 ) ≤ P ≤ φ( f 2 )

for all f 1 in L1 , f 2 in L2 . (v) If ψ = φ|R(A) , then ψ is a unital contraction, but ψ˜ = φ is not contractive. (vi) ψ has no contractive extension to S, and so none to C(∂A) either. (vii) ψ is not 2-contractive. Part (vi), in particular, shows that a direct generalization of the Hahn–Banach extension theorem to operator-valued mappings fails even when the domain is commutative. To verify the claim, it will sufﬁce to construct an element of S that is not partitionable. Fix a small positive number δ > 0 and set j = 1, 2, X j = {R j eit : δ ≤ t ≤ π − δ}, Y j = {R j eit : π + δ ≤ t ≤ 2π − δ}. Deﬁne a continuous function f on ∂A by setting f (X 1 ) = 1, f (Y1 ) = −1, f (X 2 ) = i, f (Y2 ) = −i, and extending f linearly on the remaining arcs, so that f = 1. It is easy to check that s1 ( f ) = s2 ( f ), so that f is in S. Suppose that we were given scalars λ1 , . . . , λn , |λ j | ≤ 1, j = 1, . . . , n, and positive functions p1 , . . . , pn in S, summing to less than 1, such that f is not merely approximated by the sum, but in fact λi pi . f = i

Since f (X 1 ) = 1 and |λ j | ≤ 1, we must have that some subset of the λ j ’s is exactly equal to 1, which, after reordering, we may take to be λ1 , . . . , λk . It is not difﬁcult to see that the above equations imply that p1 (x) + · · · + pk (x) = 1 for all x in X 1 . Similarly, after reordering the remaining λ j ’s, we ﬁnd a set λk+1 , . . . , λm that are all equal to −1, and pk+1 (x) + · · · + pm (x) = 1, for all x in Y1 . But this implies that s1 ( p1 + · · · + pm ) ≥ 1 − 4δ,

82

Chapter 6. Completely Positive Maps into Mn

and since s1 and s2 agree on S, s2 ( pm+1 + · · · + pn ) ≤ 4δ. We have that Im( f ) = Im(λm+1 ) pm+1 + · · · + Im(λn ) pn , so 1 − 4δ ≤ s1 (|Im( f )|) ≤ s2 ( pm+1 + · · · + pn ) ≤ 4δ, a clear contradiction for δ < 1/8. Thus, we see that no sum of the above form could actually equal f . An analogous, but somewhat more detailed, argument shows that for the same δ and ε sufﬁciently small, no sum of the above form can approximate f to within ε. We leave it to the reader to verify this claim.

Notes The correspondence between linear maps from a space into Mn and linear functionals on the tensor product of that space with Mn is a recurring theme in this subject (see for example, [46], [47], [134], [223]). Our presentation owes a great deal to [6] and [218]. The study of S + ⊗ Mn+ is adapted from [46], where it is used to study positive maps between matrix algebras. In contrast to Proposition 4.7, a number of unanswered questions and surprising examples arise in the study of positive maps between matrix algebras ([46], [222], and [244]). For example, positive maps between Mn and Mk are characterized by linear functionals on Mn (Mk ), which are positive on Mn+ ⊗ Mk+ . A decent characterization of the matrices in Mn (Mk ) = Mnk that belong to this cone is still not available. See [46] and [245] for an introduction to this topic. Example 6.11 is adapted from a similar example in [77]. The partition-of-unity techniques originate there also.

Exercises ∗

6.1 Let A be any unital C -algebra. Give an example of a linear functional s: Mn (A) → C such that s is unital and positive, but such that the associated linear map, φs : A → Mn , has norm n. (Hint: First consider the case A = C.) 6.2 Let φ: S → Mn be positive and set φ(1) = P. Let Q be the projection onto the range of P, and let R be positive with (I − Q)R = 0, R P R = Q. Let ψ: S → Mn be any positive, unital map, and set φ (a) = Rφ(a)R + (I − Q)ψ(a)(I − Q).

Exercises

6.3

6.4 6.5

6.6

83

(i) Show that φ is a unital, positive map. (ii) Show that if (ai, j ) is positive in Mk (S), but φk ((ai, j )) is not positive, then φk ((ai, j )) is not positive either. (iii) Deduce the equivalence of (i) and (ii) in Theorem 6.6. Assume that the equivalent conditions of Theorem 6.9 are not met. Show that then there always exists a unital, positive map φ: S → M2 that is not contractive. Verify the existence of a φ: S → M2 with properties (i)–(vii) of Example 6.11. Let S be an operator system. Prove that the following are equivalent: (i) For every C ∗ -algebra B, every positive φ: S → B is n-positive. (ii) S + ⊗ Mn+ is dense in Mn (S)+ . Use Corollary 6.7 to give an alternate proof of the fact that every positive map with domain C(X ) is completely positive.

Chapter 7 Arveson’s Extension Theorems

In this chapter we extend the results of Chapter 6 from ﬁnite-dimensional ranges Mn to maps with range B(H). We then develop the immediate applications of the extension theorems to dilation theory. We begin with some observations of a general functional-analytic nature. Let X and Y be Banach spaces, let Y ∗ denote the dual of Y , and let B(X, Y ∗ ) denote the bounded linear transformations of X into Y ∗ . We wish to construct a Banach space such that B(X, Y ∗ ) is isometrically isomorphic to its dual. This will allow us to endow B(X, Y ∗ ) with a weak∗ topology. Fix vectors x in X and y in Y , and deﬁne a linear functional x ⊗ y on B(X, Y ∗ ) by x ⊗ y(L) = L(x)(y). Since |x ⊗ y(L)| ≤ L · x · y, we see that x ⊗ y is in B(X, Y ∗ )∗ with x ⊗ y ≤ xy. In fact, x ⊗ y = xy (Exercise 7.1). It is not difﬁcult to check that the above deﬁnition is bilinear, i.e., x ⊗ (y1 + y2 ) = x ⊗ y1 + x ⊗ y2 , (x1 + x2 ) ⊗ y = x1 ⊗ y + x2 ⊗ y, and (λx) ⊗ y = x ⊗ (λy) = λ(x ⊗ y) for λ ∈ C. We let Z denote the closed linear span in B(X, Y ∗ )∗ of these elementary tensors. Actually, Z can be identiﬁed as the completion of X ⊗ Y with respect to a cross-norm (Exercise 7.1), but we shall not need that fact here. Lemma 7.1. B(X, Y ∗ ) is isometrically isomorphic to Z ∗ with the duality given by

L , x ⊗ y = x ⊗ y(L). Proof. It is straightforward to verify that the above pairing deﬁnes an isometric isomorphism of B(X, Y ∗ ) into Z ∗ . To see that it is onto, ﬁx f ∈ Z ∗ , and for each x, deﬁne a linear map, f x : Y → C, via f x (y) = f (x ⊗ y). Since | f x (y)| ≤ f xy, we have that f x ∈ Y ∗ . It is straightforward to verify 84

Chapter 7. Arveson’s Extension Theorems

85

that if we set L(x) = f x , then L: X → Y ∗ is linear, and that L is bounded with L ≤ f . Thus, L ∈ B(X, Y ∗ ), and clearly under our correspondence L → f , which completes the proof. We call the weak∗ topology that is induced on B(X, Y ∗ ) by this identiﬁcation the BW topology (for bounded weak). The following lemma explains the name. Lemma 7.2. Let {L λ } be a bounded net in B(X, Y ∗ ). Then L λ converges to L in the BW topology if and only if L λ (x) converges weakly to L(x) for all x in X. Proof. If L λ converges to L in the BW topology, then L λ (x)(y) = L λ , (x ⊗ y) → L , (x ⊗ y) = L(x)(y) for all y in Y . Thus, L λ (x) converges weakly to L(x) for all x. Conversely, if L λ (x) converges weakly to L(x) for all x, then L λ , (x ⊗ y) converges to L , (x ⊗ y) for all x and y and hence on the linear span of the elementary tensors. But since the net is bounded, this implies that it converges on the closed linear span. If H is a Hilbert space, then B(H) = B(H, H) is the dual of a Banach space by Lemma 7.1. This Banach space can also be identiﬁed with the space of ultraweakly continuous linear functionals or with the trace-class operators (TC) on H with trace norm, T 1 = tr (|T |), where tr (·) denotes the trace [68, Theorem I.4.5]. We prefer to focus on the duality with the trace class operators. Under this duality, an operator A ∈ B(H) is identiﬁed with the linear functional Tr(AT ) for T ∈ TC. If h, k are in H, let Rh,k denote the elementary rank-one operator on H given by Rh,k (x) = x, kh. The linear span of these operators is dense in TC in the trace norm (Exercise 7.2), and for A ∈ B(H), Tr(A Rh,k ) = Ah, k. Proposition 7.3. Let X be a Banach space, and let H be a Hilbert space. Then a bounded net {L λ } in B(X, B(H)) converges in the BW topology to L if and only if L λ (x)h, k converges to L(x)h, k for all h, k in H and x in X . Proof. We have that {L λ } converges in the BW topology to L if and only if Tr(L λ (x)T ) → Tr(L(x)T ) for all T ∈ TC and x ∈ X . But again, since the net is bounded, it is enough to consider T = Rh,k .

86

Chapter 7. Arveson’s Extension Theorems

Let A be a C ∗ -algebra, S an operator system, M a subspace. We make the following deﬁnitions: Br (M, H) = {L ∈ B(M, B(H)): L ≤ r }, CBr (M, H) = {L ∈ B(M, B(H)): Lcb ≤ r }, CPr (S, H) = {L ∈ B(S, B(H)): L is completely positive, L ≤ r }, CP(S, H; P) = {L ∈ B(S, B(H)): L is completely positive, L(1) = P}. Theorem 7.4. Let A be a C ∗ -algebra, let S be a closed operator system, and let M be a closed subspace. Then each of the four above sets is compact in the BW topology. Proof. Since the BW topology is a weak∗ topology, the set Br (M, H) is compact by the Banach–Alaoglu theorem. Since the remaining sets are subsets of this set, it is enough to show that they are closed. We argue for CBr (M, H); the rest are similar. Let {L λ } be a net in CBr (M, H), and suppose {L λ } converges to L. If (ai, j ) is in Mn (M), and x = x1 ⊕ · · · ⊕ xn , y = y1 ⊕ · · · ⊕ yn are in H ⊕ · · · ⊕ H, then (L(ai, j ))x, y = limλ (L λ (ai, j ))x, y. Hence, (L(ai, j )) ≤ r · (ai, j ) for all n, and so Lcb ≤ r . We’re now in a position to prove the main result of the chapter. Theorem 7.5 (Arveson’s extension theorem). Let A be a C ∗ -algebra, S an operator system contained in A, and φ: S → B(H) a completely positive map. Then there exists a completely positive map, ψ: A → B(H), extending φ. Proof. Let F be a ﬁnite-dimensional subspace of H, and let φF : S → B(F ) be the compression of φ to F, i.e., φF (a) = PF φ(a)|F , where PF is the projection onto F. Since B(F ) is isomorphic to Mn for some n, by Theorem 6.2 there exists a completely positive map ψF : A → B(F ) extending φF . Let ψF : A → B(H) be deﬁned by setting ψF (a) equal to ψF (a) on F and extending it to be 0 on F ⊥. The set of ﬁnite-dimensional subspaces of H is a directed set under inclusion, and so {ψF } is a net in CPr (A, H) where r = φ. Since this latter set is compact, we may choose a subnet which converges to some element ψ ∈ CPr (A, H). We claim that ψ is the desired extension. Indeed, if a ∈ S and x, y are in H. Let F be the space spanned by x and y. Then for any F1 ⊇ F, φ(a)x, y =

ψF 1 (a)x, y, and since the set of such F1 is coﬁnal, we have that φ(a)x, y =

ψ(a)x, y. This completes the proof of the theorem.

Chapter 7. Arveson’s Extension Theorems

87

Corollary 7.6 (Arveson). Let A be a C ∗ -algebra, M a subspace with 1 ∈ M, and φ: M → B(H) a unital, complete contraction. Then there exists a completely positive map ψ: A → B(H ) extending φ. We’ve seen earlier that positive maps need not have positive extensions (Example 2.13) and that unital, contractive maps need not have contractive extensions [Example 6.11(vi)] even when the range is ﬁnite-dimensional. These facts make the above results all the more striking. A C ∗ -algebra B is called injective if for every C ∗ -algebra A and operator system S contained in A, every completely positive map φ: S → B can be extended to a completely positive map on all of A. Thus, Theorem 7.5 is the assertion that B(H) is injective. Exercise 7.5 gives an elementary characterization of injective C ∗ -algebras. Corollary 7.6 is the basis for a general dilation theory. Let B be a unital C ∗ -algebra, and let A be a subalgebra (not necessarily ∗-closed) with 1 ∈ A. We shall call A an operator algebra. A unital homomorphism ρ: A → B(H) is said to have a B-dilation if there exists a Hilbert space K containing H and a unital ∗-homomorphism π: B → B(K) such that ρ(a) = PH π(a)|H for all a in A. This deﬁnition is motivated in part by the theory of normal ∂ X -dilations. Recall that if T ∈ B(H), then a compact set X is a spectral set for X provided that the homomorphism ρ: R(X ) → B(H) given by ρ(r ) = r (T ) is contractive. It is clear that T has a normal ∂ X -dilation if and only if ρ has a C(∂ X )-dilation. Corollary 7.7 (Arveson). Let A be an operator algebra contained in the C ∗ algebra B, let ρ: A → B(H) be a unital homomorphism, and let ρ: ˜ A + A∗ → B(H) be the positive extension of ρ. Then the following are equivalent: (i) ρ has a B-dilation, (ii) ρ is completely contractive, (iii) ρ˜ is completely positive. Moreover, in this case there exists a B-dilation π: B → B(K) such that π(B)H is dense in K. Proof. We have seen the equivalence of (ii) and (iii) in Chapter 3. If ρ has a B-dilation, then the map φ: B → B(H) deﬁned by φ(b) = PH π(b)|H is completely positive and extends ρ, so ρ˜ is completely positive.

88

Chapter 7. Arveson’s Extension Theorems

Conversely, if ρ˜ is completely positive, then we may extend ρ˜ to a completely positive map φ: B → B(H). The Stinespring representation of φ gives rise to the B-dilation of ρ. A minimal Stinespring representation of φ will have the property that π (B)H is dense in K. A B-dilation with the property that π(B)H is dense in K will be called a minimal B-dilation of ρ. These need not be unique (Exercise 7.3). Let T ∈ B(H), and let X be a spectral set for T . If the homomorphism ρ: R(X ) → B(H) is completely contractive, then we shall call X a complete spectral set for T . Corollary 7.8. Let T ∈ B(H), and let X be a spectral set for T . Then the following are equivalent: (i) (ii) (iii) (iv)

T has a normal ∂ X -dilation, X is a complete spectral set, ρ˜ is completely positive, ρ˜ has a positive extension to C(∂ X ).

Moreover, in this case there is a normal ∂ X -dilation N for T such that the smallest closed, reducing subspace for N containing H is K. Proof. The equivalence of (i)–(iii) is just Corollary 7.7. If ρ˜ is completely positive, then it has a (completely) positive extension to C(∂ X ) by Arveson’s extension theorem. Conversely, if ρ˜ does have a positive extension to C(∂ X ), then that extension is automatically completely positive by Theorem 3.11, and hence its restriction to S, ρ, ˜ is completely positive. If K is the smallest closed, reducing subspace for N containing H, then the representation π : C(∂ X ) → B(K) given by π (z) = N has no closed, reducing subspaces containing H. But this is equivalent to the requirement that π (C(∂ X ))H be dense in K, since this latter space is clearly reducing. A normal ∂ X -dilation of T with no closed, reducing subspace containing H is called a minimal normal ∂ X -dilation of T . Unlike the case of Sz.-Nagy’s minimal unitary dilation of a contraction (Theorem 4.3), the minimal normal ∂ X -dilation of an operator need not be unique up to unitary equivalence. Exercise 7.3 illustrates the difﬁculty. A unital homomorphism of an operator algebra A into B(H) will be called a representation of A. If A is a C ∗ -algebra, then every contractive representation of A is automatically a ∗-representation. Corollary 7.7 gives a useful

Chapter 7. Arveson’s Extension Theorems

89

characterization of the completely contractive representations, for it shows that these representations can be studied via the ∗-representation theory of B. Consequently, it is important to know when representations are completely contractive. At present we have no natural source of contractive representations that are not completely contractive. In general, such examples are hard to construct, since in some sense they are the pathological cases that are difﬁcult to ﬁt to a general theory. They can occur, however, even when the algebra is ﬁnitedimensional and the representation is on a ﬁnite-dimensional space. An example of a subalgebra of M3 and a representation of it on C3 that is not completely contractive is given in [6]. An example of considerable importance to the theory of dilations of contractions was given by Parrott [156]. Parrott’s . Example. Let U and V be contractions in B(H) such that U is unitary and U and V don’t commute. We deﬁne commuting contractions on B(H ⊕ H) by setting

0 0 , T1 = I 0

0 T2 = U

0 , 0

0 T3 = V

0 . 0

Let P(D3 ) be the algebra of polynomials in three variables z 1 , z 2 , z 3 , regarded as a subalgebra of C(T3 ), where T3 is the 3-torus. In Chapter 5 we saw an example of three commuting contractions such that the induced homomorphism of P(D3 ) was not contractive. We claim that the homomorphism ρ: P(D3 ) → B(H ⊕ H) deﬁned by ρ(z i ) = Ti , i = 1, 2, 3, is contractive but not completely contractive. To see that ρ is contractive, let p(z 1 , z 2 , z 3 ) be an arbitrary element of P(D3 ), and write p(z 1 , z 2 , z 3 ) = a0 + a1 z 1 + a2 z 2 + a3 z 3 + q(z 1 , z 2 , z 3 ), where q(z 1 , z 2 , z 3 ) contains all the higher-order terms of p. We have that

a0 ρ( p) = a1 I + a2 U + a3 V since Ti · T j = 0 for all i and j.

0 , a0

90

Chapter 7. Arveson’s Extension Theorems

Let x = x1 ⊕ x2 and y = y1 ⊕ y2 be arbitrary unit vectors in H ⊕ H, and calculate | ρ( p)x, y| = |a0 x1 , y1 + (a1 I + a2 U + a3 V )x1 , y2 + a0 x2 , y2 | ≤ |a0 |x1 y1 + (|a1 | + |a2 | + |a3 |)x1 y2 + |a0 |x2 y2 0 x1 y1 |a0 | , . = |a1 | + |a2 | + |a3 | |a0 | x2 y2 Thus, we have that

|a0 | 0 , ρ( p) ≤ |a1 | + |a2 | + |a3 | |a0 |

where the latter matrix is an element of M2 . But by Exercise 2.11, 0 |a0 | {|a0 | + (|a1 | + |a2 | + |a3 |)z + r (z)}, |a1 | + |a2 | + |a3 | |a0 | ≤ inf r where r (z) is an arbitrary polynomial whose lowest-order term is at least of degree 2, and the latter norm is the supremum norm over the unit circle. Let λ0 , λ1 , λ2 , λ3 be numbers of modulus 1 such that λ0 a0 , λ1 a1 , λ2 a2 , λ3 a3 are positive; then |a0 | 0 ρ( p) ≤ |a1 | + |a2 | + |a3 | |a0 | ≤ |a0 | + (|a1 | + |a2 | + |a3 |)z + λ0 q(λ¯ 0 λ1 z, λ¯ 0 λ2 z, λ¯ 0 λ3 z) = p(λ¯ 0 λ1 z, λ¯ 0 λ2 z, λ¯ 0 λ3 z) ≤ p(z 1 , z 2 , z 3 ), where the third and fourth norms are the supremum over T and the last norm is the supremum over T3 . Thus, ρ is contractive. Now assume that ρ is completely contractive. Consider an element of Mn (P(D3 )) of the form (ai, j z 1 + bi, j z 2 + ci, j z 3 ). Its image under ρn is   0 0 ··· 0 0 .. ..     . .    a11 I + b11 U + c11 V 0 · · · a1n I + b1n U + c1n V 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   .. ..  .. .. . .  ,  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .    0 0 ··· 0 0  .. .. .. ..     . . . . an1 I + bn1 U + cn1 V 0 · · · ann I + bnn U + cnn V 0 which is an operator acting on the direct sum of n copies of H ⊕ H.

Chapter 7. Arveson’s Extension Theorems

91

Think of this space as (H11 ⊕ H21 ) ⊕ · · · ⊕ (H1n ⊕ H2n ), where each Hi, j = H. If we reorder the spaces by writing (H11 ⊕ · · · ⊕ H1n ) ⊕ (H21 ⊕ · · · ⊕ H2n ), then the matrix of this same operator will be   .. 0. · · · · · · · · · · · · · · · 0. . 0 · · · · · · · · · · · · · · · 0 .. ..  ..  .. ..  .. .. ..  .. ..  .. .. ..  .. ..   .. .. ..   . . . .  .. . .  0 ··············· 0 .. 0 · · · · · · · · · · · · · · · 0   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .    a11 I + b11 U + c11 V . . . a1n I + b1n U + c1n V ... 0 · · · · · · · · · · · · · · · 0 . .  .. . ..    .. .. .. .. ..   . . .. ..   . ..  a1n I + bn1 U + cn1 V . . . ann I + bnn U + cnn V . .. .. .  . .. 0 · · · · · · · · · · · · · · · 0 . Since ρ is assumed to be completely contractive, the operator matrix in the (2,1) block of the above matrix has norm bounded by ρn ((ai, j z 1 + bi, j z 2 + ci, j z 3 )) ≤ sup{(ai, j z 1 + bi, j z 2 + ci, j z 3 ) : |z 1 | = |z 2 | = |z 3 | = 1}. But this latter supremum is equal to sup{(ai, j + bi, j w1 + ci, j w2 ): |w1 | = |w2 | = 1}. Let M be the subspace of C(T2 ) spanned by 1, w1 , and w2 , where w1 and w2 are the coordinate functions. The above discussion shows that if ρ is completely contractive, then the map ψ: M → B(H) deﬁned by ψ(1) = I, ψ(w1 ) = U , and ψ(w2 ) = V is completely contractive. By Arveson’s extension theorem, we may extend ψ to a completely positive map on C(T2 ) and then obtain a Stinespring representation (π, S, K) with S an isometry and π: C(T2 ) → B(K) a ∗-homomorphism. Identifying SH and H, we have that ψ(·) = PH π(·)|H . But since π(w1 ) and U = ψ(w1 ) are both unitary, this implies that H reduces π(w1 ). Finally, since π(w1 ) and π(w2 ) commute, and H reduces π (w1 ), this implies that ψ(w1 ) commutes with ψ(w2 ) – that is, that U commutes with V , a contradiction. Therefore, ρ is not completely contractive. If one chooses U and V to both be unitaries, then it is possible to show that ρ is not even 4-contractive (Exercise 7.9).

92

Chapter 7. Arveson’s Extension Theorems

Another remark to be made is that there exist U and V in M2 satisfying the above conditions, and hence there are three commuting contractions in M4 for which ρ is a contraction but not a complete contraction. Parrott’s example also shows that we must be extremely careful about what we mean by the expression “operator algebra.” Consider any ∗-isomorphism π: C(T3 ) → B(K) and let Ai = π(z i ) ⊕ Ti , i = 1, 2, 3, where the Ti ’s are the operators given above. For any polynomial p in three variables we have that p(A1 , A2 , A3 ) = π ( p) ⊕ p(T1 , T2 , T3 ) and hence p(A1 , A2 , A3 ) = p∞ . Thus, the algebra A ⊆ B(K ⊕ H ⊕ H ) generated by A1 , A2 , and A3 and the algebra P(D3 ) ⊆ C(T3 ) are isometrically isomorphic via the map γ : P(D3 ) → A, γ ( p) = p(A1 , A2 , A3 ). Hence, we could regard them as the “same” Banach algebra. However, we should not regard them as the “same” operator algebra, because, by the above calculations, we know that γ is not a complete isometry. Thus, the norms on Mn (A) differ from the norms on Mn (P(D3 )). In fact, since ρ is not 4-contractive, we have that γ4 : M4 (P(D3 )) → M4 (A) is not an isometry. Thus, P(D3 ) and A are distinct operator algebras, since in this book we are concerned with their matrix norm structures. This subtlety does not arise with C ∗ -algebras. The discussion in Chapter 1 shows that if γ : A → B is a ∗-isomorphism between two C ∗ -algebras, then γn : Mn (A) → Mn (B) is also a ∗-isomorphism for all n. Hence every isometric isomorphism between C ∗ -algebras is automatically a completely isometric isomorphism. Now that we’ve seen an example of a commutative operator algebra for which every contractive representation is not necessarily completely contractive, we would like to present an example in the opposite extreme. That is, some highly noncommutative operator algebras for which every contractive representation is completely contractive. In general, it is hard to test if a homomorphism is contractive. Thus, the following theorem is quite nice. Proposition 7.9 (McAsey–Muhly). Let A be the algebra of upper triangular matrices in Mn , i.e., A = {(ai, j ): ai, j = 0 for i > j}, and let {E i, j } be the standard matrix units for Mn . Then every representation ρ of A with ρ(E i j ) ≤ 1 for i ≤ j is completely contractive. Proof. Let ρ: A → B(H) be a representation. We deﬁne a map φ: Mn → B(H) by φ(E i j ) = ρ(E i j ), i ≤ j, and φ(E i j ) = ρ(E ji )∗ , i > j. We must prove that φ is completely positive.

Chapter 7. Arveson’s Extension Theorems

93

Since ρ is unital and a homomorphism, {ρ(E i,i )} will be projections that sum to the identity. Since ρ(E ii ) ≤ 1, they will be orthogonal projections. Let Hi be the space that ρ(E i,i ) projects onto, so H = H1 ⊕ · · · ⊕ Hn . Since ρ(E i, j ) = ρ(E i,i )ρ(E i, j )ρ(E j, j ), there will exist operators Ti, j : H j → Hi such that the operator matrix of ρ(E i, j ) relative to the above decomposition of H is Ti, j in the (i, j)th position and 0 elsewhere. Relative to this decomposition we have that ρ((ai j )) = (ai j Ti j ). Set Ti = Ti,i+1 , and note that for i < j, Ti, j = Ti · Ti+1 · · · T j . By Theorem 3.14, to prove that φ is completely positive, it is sufﬁcient to prove that (φ(E i, j )) is a positive operator on the direct sum of n copies of H, H1 ⊕ · · · ⊕ Hn . Decomposing each Hi as H1i ⊕ · · · ⊕ Hni , we have that (φ(E i, j )) is represented as an n 2 × n 2 operator matrix with each n × n block having only one nonzero entry. Reorder the subspaces so that H11 , H22 , . . . , Hnn are listed ﬁrst and the remaining spaces occur in any order. The operator matrix for (φ(E i, j )) with respect to this reordering is (Ti, j ) in the ﬁrst n × n entries and 0 elsewhere. Thus, it will be sufﬁcient to prove that (Ti, j ) is positive. Set 

0 T1  0 0 . . R=  .. . . 0 ...

0 T2 .. .

... .. . .. . .. .

...

...

 0 ..  .   0  ;  Tn−1  0

then R ≤ 1, R n+1 = 0, and (Ti, j ) = (I − R)−1 + (I − R ∗ )−1 − I . Conse quently, as in the proof of Theorem 2.6, (Ti, j ) is positive. If we examine the Stinespring representation of ρ, then it can be seen to give a factorization of the operator (Ti, j ). Indeed, if ρ(A) = V ∗ π (A)V for some ∗-homomorphism π: Mn → B(K) and isometry V : H → K, then we know that up to a unitary equivalence we have that K = L(n) for some Hilbert space L and π(E i, j ) is the operator matrix that is IL in the (i, j) entry and 0 elsewhere. Writing V : H → L(n) as V h = (V1 h, . . . , Vn h) where Vi : H → L, we have that ρ(E i, j ) = Vi∗ V j . Since so many positive maps on Mn are not completely positive, it is somewhat surprising that every map induced by a representation of A is completely positive.

94

Chapter 7. Arveson’s Extension Theorems Notes

The BW topology was introduced in [6], where the proofs of the extension theorems also appear, as well as some material on B-dilations. Parrott [156] originally proved that the set of operators in his example did not have a set of commuting unitary dilations, i.e., that ρ(z 1 , z 2 , z 3 ) → ρ(T1 , T2 , T3 ) does not have a C(T3 ) dilation, by examining spatial relations among the unitaries. By Corollary 7.7, this is equivalent to the fact that Parrott’s example yields a contractive homomorphism that is not completely contractive. We’ve avoided Parrott’s spatial proof in an attempt to cast a different perspective. It is interesting to note that when V is also unitary, a direct proof of the noncomplete contractivity of the homomorphism simpliﬁes (Exercise 7.9), as does the spatial proof [231]. We do not know if the homomorphism is not 2-contractive for arbitrary V . Let us reinterpret the material on commuting contractions in light of Corollary 7.7. Ando’s theorem [5] shows that every pair of commuting contractions induces a completely contractive homomorphism of the bidisk algebra P(D2 ). So, in particular, every contractive homomorphism of that algebra is completely contractive. Parrott’s example shows that P(D3 ) differs in this respect; contractive homomorphisms need not be completely contractive. The internal properties of the algebra P(D3 ) that lead to this difference do not seem to be understood. Neither do the internal properties of the bidisk algebra. For example, it is not known if every unital contraction (including nonhomomorphisms) on P(D2 ) is completely contractive. This is the case for the disk algebra, since it is a Dirichlet algebra. We shall see some additional explanation of the difference between the two- and three-variable situations in Chapter 18, but our understanding of this situation is still not satisfactory. The results of Crabbe and Davies [69] and Varopoulos [236] examined in Chapter 5, that not every set of n ≥ 3 commuting contractions need satisfy von Neumann’s inequality, reﬂect a different property of the polydisk algebras than Parrott’s example. In Chapter 18, we will formalize this difference a bit more. Gaspar and Racz [104] have shown that every set of cyclically commuting contractions has a set of cyclically commuting, unitary dilations. If given a C ∗ -algebraic interpretation, their result should add some additional insights. Choi and Effros [50] prove that for a von Neumann algebra A, semidiscreteness (Exercise 7.14) is equivalent to the existence of a net of ﬁnite-rank, completely positive maps Rλ : A → A such that Rλ (x) converges to x σ -weakly for all x. Connes [67] proves that semidiscreteness and injectivity are equivalent for von Neumann algebras. Proposition 7.9 is adapted from McAsey and Muhly [143].

Exercises

95

Exercises 7.1 Let X and Y be Banach spaces, and for f ∈ X ∗ , g ∈ Y ∗ deﬁne L f,g ∈ B(X, Y ∗ ) by L f,g (x) = f (x)g. By considering these operators, prove that the map j: X ⊗ Y → B(X, Y ∗ )∗ deﬁned in this chapter has the following properties: (i) j is linear, (ii) j(x ⊗ y) = xy, (iii) j is one-to-one. Since j is one-to-one, the identiﬁcation of X ⊗ Y with j(X ⊗ Y ) endows X ⊗ Y with a norm (rather than just a seminorm). Conclude that Z of Lemma 7.1 can be identiﬁed with the completion of X ⊗ Y with respect to this norm. A norm on the tensor product of two normed spaces that satisﬁes x ⊗ y = xy is called a cross-norm. 7.2 Let Rh,k be the operator deﬁned by Rh,k (x) = x, kh. Show that Rh,k is trace-class, that Tr(A Rh,k ) = Ah, k, and that the linear span of {Rh,k : h, k ∈ H} is dense in TC in the trace norm. 7.3 Let A be an operator algebra contained in the C ∗ -algebra B, let ρ: A → B(H) be a completely contractive, unital homomorphism, and let πi : B → B(Ki ), i = 1, 2, deﬁne minimal B-dilations of ρ. Deﬁne completely positive maps φi : B → B(H) by φi (b) = PH πi (b)|H , i = 1, 2. (i) Show that there exists a unitary U : K1 → K2 with U h = h for h in H, and U ∗ π2 (b)U = π1 (b) if and only if φ1 = φ2 . Such dilations are called unitarily equivalent. (ii) Show that there is a one-to-one correspondence between unitarily equivalent, minimal B-dilations of ρ and completely positive extensions of ρ to B. (iii) Show that the set of completely positive extensions of ρ is a compact, convex set in the BW toplogy on CP(B, H). 7.4 (Extension of bimodule maps) Let A, C be C ∗ -algebras, let S be an operator system, and suppose that C ⊆ S ⊆ A. If C ⊆ B(H), then φ: S → B(H) is a C-bimodule map provided φ(c1 ac2 ) = c1 φ(a)c2 . Prove that if φ: S → B(H) is a completely positive C-bimodule map, then every completely positive extension of φ to A is also a C-bimodule map. 7.5 Let B ⊆ B(H) be a unital C ∗ -algebra. Prove that B is injective if and only if there exists a completely positive map φ: B(H) → B such that φ(b) = b for all b in B. Show that φ is necessarily a B-bimodule map. A map with the above properties is called a completely positive conditional expectation. 7.6 (Sarason) Let A ⊆ B(K) be an algebra, and let H be a subspace of K such that A → PH A|H is a homomorphism on A. Prove that the subspaces H1 =

96

7.7

7.8 7.9 7.10

7.11

7.12

7.13

Chapter 7. Arveson’s Extension Theorems span{Ah: A ∈ A, h ∈ H}− and H2 = H1 ∩ H⊥ of K both are invariant for every element of A, and that H1 ⊕ H = H2 . Conversely, show that if H1 and H2 are invariant for A and H1 ⊕ H = H2 , then compression onto H is a homomorphism on A. Such a subspace is said to be semiinvariant for A. Let Q = (qi, j ) be the element of Mn given by qi,i+1 = 1, qi, j = 0 for all other (i, j), and let N ∈ B(H) satisfy N ≤ 1, N n = 0. Prove that Q → N deﬁnes a completely contractive representation of the algebra generated by Q. Prove that B is injective if and only if Mn (B) is injective for some n. In Parrott’s example, show that if U and V are both unitaries that don’t ∗ V ].) commute, then ρ is not 4-contractive. (Hint: Consider [ UV ∗ −U (Parrott) Prove that there exist three commuting operators, T1 , T2 , and T3 , with Ti < 1 such that von Neumann’s inequality holds, but that do not have a commuting unitary dilation. Prove that for every n ≥ 3, there exist n commuting operators such that Ti < 1, i = 1, . . . , n, and von Neumann’s inequality holds, but that do not have commuting unitary dilations. Let A ⊆ B(H) be a unital C ∗ -algebra. If A is closed in the weak operator topology, then A is called a von Neumann algebra. Prove that every von Neumann algebra has a predual. The induced weak∗ topology on A is called the σ -weak topology. Let A be a von Neumann algebra. If there exists a net of σ -weakly continuous, completely positive maps, φλ : A → Mn λ , ψλ : Mn λ → A, with φλ (1) = 1, ψλ (1) = 1, such that ψλ ◦ φλ (a) → a, σ -weakly for all a in A, then A is called semidiscrete. Prove that every semidiscrete von Neumann algebra is injective.

Chapter 8 Completely Bounded Maps

In this chapter, we extend many of the results obtained in previous chapters concerning completely positive maps to the completely bounded maps. Our main technique is to realize completely bounded maps as the off-diagonal corners of completely positive maps. Let us consider for a moment Mm (Mn (A)). For a C ∗ -algebra A, a typical element of this algebra is of the form A = (Ai, j )i,m j=1 , where each Ai, j is in Mn (A). Thus, Ai, j = (ai, j,k, )nk,=1 , with ai, j,k, in A. Setting Bk, = (ai, j,k, )i,m j=1 , we obtain an element of Mm (A), and thus B = (Bk, )nk,=1 is in Mn (Mm (A)). Now, Mm (Mn (A)) and Mn (Mm (A)) are both isomorphic to Mnm (A) by just deleting the extra parentheses. With these identiﬁcations, A and B are unitarily equivalent elements of Mnm (A); in fact, the unitary is just a permutation matrix. To see this, note that if we regard A as an element of Mmn (A), say A = (cs,t )mn s,t=1 , then cs,t = ai, j,k, , where s = n(i − 1) + k, t = n( j − 1) + ; while if we regard B as an element of Mmn (A), say B = (ds,t )mn s,t=1 , then ds,t = ai, j,k, , where s = m(k − 1) + i, t = m( − 1) + j. Since the above operation for passing from Mm (Mn (A)) to Mn (Mm (A)) is just a permutation, it is a ∗-isomorphism. We shall refer to this ∗-isomorphism as the canonical shufﬂe. It is important to note that since the canonical shufﬂe is a ∗-isomorphism, it preserves norm and positivity. Note that we’ve encountered this permutation earlier in Parrott’s example, in the proof of Proposition 7.9, and in our discussion of the Kronecker product in Chapter 3. It is also useful to understand this canonical shufﬂe in the tensor notation. Let {E i, j }i,m j=1 and {Fk, }nk,=1 denote the standard matrix units for Mm and Mn , respectively. Our element A of Mm (Mn (A)) ∼ = (A ⊗ Mn ) ⊗ Mm m is just A = i, j=1 Ai, j ⊗ E i, j , where each Ai, j is in A ⊗ Mn and has the form Ai, j = nk,=1 ai, j,k, ⊗ Fk, . Thus, A = i,m j=1 nk,=1 ai, j,k, ⊗ Fk, ⊗ E i, j . m On the other hand, Bk, = i, j=1 ai, j,k, ⊗ E i, j , so that B = nk,=1 Bk, ⊗

97

98

Chapter 8. Completely Bounded Maps

∼ Fk, = nk,=1 i,m j=1 ai, j,k, ⊗ E i, j ⊗ Fk, , which lies in (A ⊗ Mm ) ⊗ Mn = Mn (Mm (A)). When we regard A and B as elements of Mmn (A), then we see that B is the image of A under the following string of isomorphisms: Mmn (A) ∼ = Mm (Mn (A)) ∼ = (A ⊗ Mn ) ⊗ Mm ∼ = A ⊗ (Mm ⊗ Mn ) ∼ = = A ⊗ (Mn ⊗ Mm ) ∼ (A ⊗ Mm ) ⊗ Mn ∼ = Mn (Mm (A)) ∼ = Mmn (A). Most of these isomorphisms are so canonical that it is customary to ignore them. Setting A = C, it is interesting to note that when we identify Mmn = Mm (Mn ) = Mn ⊗ Mm and Mmn = Mn (Mm ) = Mm ⊗ Mn , then it is only the isomorphism between Mm ⊗ Mn and Mn ⊗ Mm that leads to the permutation in Mmn . The following lemma is central to this chapter and introduces the off-diagonal technique. Lemma 8.1. Let A, B be C ∗ -algebras with unit 1, let M be an operator space in A, and let φ: M → B. Deﬁne an operator system SM ⊆ M2 (A) by λ1 a : λ, µ ∈ C, a, b, ∈ M , SM = b∗ µ1 and : SM → M2 (B) via λ1 φ(a) λ1 a = . φ(b)∗ µ1 b∗ µ1 If φ is completely contractive, then is completely positive. Proof. Let (Si, j ) be in Mn (SM ), say λi, j Si, j = ∗ bi, j

ai, j µi, j

.

Since Mn (SM ) is a subspace of Mn (M2 (A)), if we perform the canonical shufﬂe, then (Si, j ) becomes an element of M2 (Mn (A)). Indeed, if we set H = (λi, j ), A = (ai, j ), B = (b j,i ), K = (µi, j ), then the image of (Si, j ) under the canonical shufﬂe is H A . (8.1) B∗ K Similarly, after the canonical shufﬂe, n ((Si, j )) becomes H φn (A) . φn (B)∗ K

(8.2)

Thus, to prove that is completely positive, it is sufﬁcient to prove that for all n, if (8.1) is positive, then (8.2) is positive.

Chapter 8. Completely Bounded Maps

99

Now, if (8.1) is positive, then A = B and H and K must be positive. Fix ε > 0, and set Hε = H + ε I, K ε = K + ε I , so that Hε and K ε are positive and invertible. We have that −1/2 −1/2 −1/2 Hε I Hε AK ε 0 = −1/2 −1/2 −1/2 0 Kε K ε A∗ Hε I −1/2 Hε 0 Hε A × −1/2 A∗ K ε 0 Kε is positive, and consequently by Lemma 3.1, −1/2 H AK ε−1/2 ≤ 1. ε −1/2

Also, φn (Hε Hε

−1/2

−1/2

−1/2

) = Hε φn (A)K ε (Exercise 8.1), and so 1/2 Hε 0 φn (A) = 1/2 φn (A)∗ Kε 0 Kε

−1/2 I φn H −1/2 AK −1/2 0 Hε × .

∗ 1/2 I φn H −1/2 AK −1/2 0 Kε AK ε

But, since φ is completely contractive, φn (H −1/2 AK −1/2 ) ≤ 1, and so by another application of Lemma 3.1, the middle matrix on the right of the above equation is positive. Thus, the left hand side is positive for all ε, and so (8.1) is positive. Consequently, is completely positive, which completes the proof. We can now prove a Hahn–Banach-type extension theorem for completely bounded maps. Theorem 8.2 (Wittstock’s extension theorem). Let A be a unital C ∗ -algebra, M a subspace of A, and let φ: M → B(H) be completely bounded. Then there exists a completely bounded map ψ: A → B(H) that extends φ, with φcb = ψcb . Proof. We may assume without loss of generality that φcb = 1. Let SM and be as in Lemma 8.1. Since is completely positive, by Arveson’s extension theorem there exists a completely positive : M2 (A) → M2 (B(H)) = B(H ⊕ H) extending .

100

Chapter 8. Completely Bounded Maps

Deﬁne ψ via

0 a ∗ ψ(a) = . 0 0 ∗ ∗

Clearly, ψ is linear, and since extends , ψ extends φ. Also, since is unital, 0 a ≤ · a ≤ a, ψ(a) ≤ 0 0 and so ψ is contractive. To see that ψ is completely contractive, let A = (ai, j ) be in Mn (A); then

n n 0 ai, j ∗ ψ(ai, j ) n . = 0 0 ∗ ∗ i, j=1 i, j=1 After performing the canonical shufﬂe, the right hand side becomes ∗ ψn (A) , ∗ ∗ where the asterisk now indicates an n × n rather than a 1 × 1 entry. Thus, by the same inequality as above, n 0 ai, j . ψn (A) ≤ 0 0 i, j=1 But after a canonical shufﬂe, n 0 ai, j 0 A becomes . 0 0 0 0 i, j=1 Thus, ψn (A) ≤ A, which is what we needed to prove.

If M is an operator space in the C ∗ -algebra A, B is another C ∗ -algebra, and φ : M → B is a linear map, we set M∗ = {a: a ∗ ∈ M} and deﬁne a linear map φ ∗ : M∗ → B,

via φ ∗ (a) = φ(a ∗ )∗ .

When M = M∗ , we set Re φ = (φ + φ ∗ )/2,

Im φ = (φ − φ ∗ )/2i,

so that Re φ, Im φ are self-adjoint, linear maps with φ = Re φ + i Im φ. We note that some care is needed, since, in general, (Re φ)(a) = Re(φ(a)), but they are equal when a = a ∗ .

Chapter 8. Completely Bounded Maps

101

The above lemma also yields the following decomposition theorem for completely bounded maps. Theorem 8.3. Let A be a C ∗ -algebra with unit, and let φ: A → B(H) be completely bounded. Then there exist completely positive maps φi : A → B(H) with φi cb = φcb , i = 1, 2, such that the map : M2 (A) → B(H ⊕ H) given by φ1 (a) φ(b) a b = c d φ ∗ (c) φ2 (d) is completely positive. Moreover, if φcb ≤ 1, then we may take φ1 (1) = φ2 (1) = IH . Proof. Clearly, we may assume that φcb = 1. Applying Lemma 8.1 with M = A, we obtain a completely positive map : SA → B(H ⊕ H) where SA ⊆ M2 (A). By Arveson’s extension theorem, extends to a completely positive map on all of M2 (A), which we still denote by . Since [ 0c b0 ] is in SA , by the deﬁnition of , 0 φ(b) 0 φ(b) 0 b = . = c 0 0 φ ∗ (c) 0 φ(c∗ )∗ Now, let p be positive, p ≤ 1; then since p 0 1 0 ≤ , 0 0 0 0 we have 0 0 p ≤ 0 0 0 1 ≤ 0

0 0 0 1 0 = . 0 0 0

A straightforward calculation shows that these inequalities taken together imply that

p 0 0 0

∗ 0 = . 0 0

102

Chapter 8. Completely Bounded Maps

Since A is the span of its positive elements, there must exist a linear map φ1 : A → B(H) such that φ1 (a) 0 a 0 . = 0 0 0 0 By an argument similar to the proof in Theorem 8.2, that ψ is completely contractive, one obtains that φ1 is completely positive. By analogous arguments, one obtains a completely positive map φ2 : A → B(H) satisfying 0 0 0 0 . = 0 d 0 φ2 (d) Thus, we see that any completely positive from SA to M2 (A) has the desired form, a b 0 b a = + c d c 0 0

map obtained via extension 0 d

φ1 (a) φ(b) , = φ ∗ (c) φ2 (d)

for some φ1 and φ2 . Since φ1 (1) = φ2 (1) = 1, we have that φ1 cb = φ2 cb = φcb . This com pletes the proof of the theorem. An alternative way to prove the above theorem is to use the fact that ﬁxes the C -subalgebra, C ⊕ C, consisting of scalar diagonal matrices, and consequently by Corollary 3.19 is necessarily a bimodule map over this algebra. An extension of this idea is included in Exercise 8.6. The above decomposition leads readily to a generalization of Stinespring’s representation. ∗

Theorem 8.4. Let A be a C ∗ -algebra with unit, and let φ: A → B(H) be a completely bounded map. Then there exists a Hilbert space K, a ∗-homomorphism π : A → B(K), and bounded operators Vi : H → K, i = 1, 2, with φcb = V1 · V2 such that φ(a) = V1∗ π(a)V2 for all a ∈ A. Moreover, if φcb = 1, then V1 and V2 may be taken to be isometries. Proof. Clearly, we may assume that φcb = 1. Let φ1 , φ2 , and be as in Theorem 8.3. Let (π1 , V, K1 ) be a minimal Stinespring representation for ,

Chapter 8. Completely Bounded Maps

103

and note that since is unital, V may be taken to be an isometry and π1 to be unital. Since M2 (A) contains a copy of M2 , the Hilbert space K1 may be decomposed as K1 = K ⊕ K in such a way that the ∗-homomorphism π1 : M2 (A) → B(K ⊕ K) has the form a b π(a) π (b) π1 = , c d π (c) π(d) where π : A → B(K) is a unital ∗-homomorphism (Exercise 8.3). Thus, we have that V : H ⊕ H → K ⊕ K is an isometry, and φ1 (a) φ(b) ∗ π(a) π (b) =V V. φ ∗ (c) φ2 (d) π (c) π(d) For h in H,

h φ (1) = 1 0 0 1 = V∗ K 0

0 h h ∗ π(1) 0 V =V 0 0 0 0 0 0 h V , 0 0

and since V is an isometry, we must have that h ∗ V = . 0 0 Thus, there is a linear map V1 : H → K such that V1 h h , V = 0 0 and V1 must also be an isometry. Similarly, there exists V2 : H → K such that 0 0 V . = h V2 h Consequently,

φ1 (a) φ(b) π(a) π (b) = V∗ V π(c) π(d) φ ∗ (c) φ2 (d) V1∗ π(a)V1 V1∗ π (b)V2 , = V2∗ π(c)V1 V2∗ π (d)V2

which completes the proof of the theorem.

104

Chapter 8. Completely Bounded Maps

Unlike the Stinespring representation of a completely positive map, there are no extra conditions known to impose on the above representation of a completely bounded map that will make it unique up to unitary equivalence. Given , one can always take its minimal (and hence unique) Stinespring representation, but unfortunately is not uniquely determined by φ. Recall that φ uniquely determines a completely positive map on the operator system λ a : λ, µ ∈ C, a, b ∈ A ⊆ M2 (A) S= b∗ µ and that is obtained by extending this completely positive map from S to M2 (A). Thus, in a certain sense the nonuniqueness of the representation is parameterized by the set of possible completely positive extensions. It is possible to impose conditions on the above representation such that it is unique up to conjugation by a (generally unbounded) closed, densely deﬁned similarity [171]. Theorem 8.5 (Wittstock’s decomposition theorem). Let A be a C ∗ -algebra with unit, and let φ: A → B(H) be completely bounded. Then there exists a completely positive map ψ: A → B(H) with ψcb ≤ φcb such that ψ ± Re(φ) and ψ ± Im(φ) are all completely positive. In particular, the completely bounded maps are the linear span of the completely positive maps. Proof. Let φ(a) = V1∗ π(a)V2 be as in Theorem 8.4 with V1 = V2 = φcb , and set ψ(a) = (V1∗ π(a)V1 + V2∗ π(a)V2 )/2, so that ψ is completely positive and ψcb = ψ(1) ≤ φcb . Notice that φ ∗ (a) = V2∗ π (a)V1 , so that 1/2

2ψ(a) ± 2 Re φ(a) = (V1 ± V2 )∗ π (a)(V1 ± V2 ), 2ψ(a) ± 2 Im φ(a) = (V1 ∓ i V2 )∗ π (a)(V1 ∓ i V2 ), and each of these four maps is completely positive. For the last statement, note that 2φ = ((ψ + Re φ) − (ψ − Re φ)) + i((ψ + Im φ) − (ψ − Im φ)) is a decomposition of φ into the span of four completely positive maps.

Operator-Valued Measures Let X be a compact Hausdorff space, let E be a bounded, regular, operatorvalued measure on X , and let φ: C(X ) → B(H) be the bounded, linear map

Operator-Valued Measures

105

associated with E by integration as described in Chapter 4. We call E completely bounded when φ is completely bounded. By Wittstock’s decomposition theorem, E is completely bounded if and only if it can be expressed as a linear combination of positive operator-valued measures. When H is one-dimensional, that is, when E is a bounded, regular, complexvalued measure on X , then the associated bounded, linear map φ: C(X ) → C is automatically completely bounded by Proposition 3.8. Wittstock’s decomposition becomes the statement that every complex measure is the span of four positive measures. If we let ψ be deﬁned by integration against the total variation measure |E| associated with E, then ψ ± Re φ and ψ ± Im φ are completely positive maps, and ψcb = ψ(1) = |E|(X ) = φ = φcb . Thus, we see that in some sense, the ψ of Wittstock’s decomposition can be thought of as a total variation or an “absolute value” of the completely bounded map φ. This analogy is pursued in [139] and [171]. In a closely related development, Loebl [139] deﬁnes a self-adjoint map φ: C(X ) → B(H) to have ﬁnite total variation if sup |φ( f i )|: f i is a partition of unity is ﬁnite, and proves that every such map decomposes as φ = φ1 − φ2 with φ1 and φ2 (completely) positive. In particular, this shows that such maps are completely bounded. However, Hadwin [109] has shown that there are completely bounded, self-adjoint maps that are not of ﬁnite total variation. Unfortunately, there is no known analytic characterization of the completely bounded, operator-valued measures. To illustrate the difﬁculties, let us suppose for simplicity that E is self-adjoint and completely bounded. From Wittstock’s decomposition, we obtain a positive, operator-valued measure F such that F(B) ± E(B) ≥ 0 for all Borel sets B. If E and F were scalar-valued, this inequality would imply that F(B) ≥ |E(B)|, but for operators, this is far from the case. In fact, it is possible for E to be completely bounded while sup |E(Bi )|: Bi disjoint, Borel = +∞. This phenomenon was ﬁrst described by Hadwin [109], and we reproduce his example here. Let X = {xn }∞ n=1 be a countable, compact Hausdorff space, and let An , n = 1, 2, . . . , be a sequence of self-adjoint operators on H. If these operators are summable in the weak operator topology, then setting E({xn }) = An deﬁnes a self-adjoint, operator-valued measure on X .

106

Chapter 8. Completely Bounded Maps

Now, let H be a separable Hilbert space with orthonormal basis, e1 , e2 , . . . , and deﬁne An via An f =

1 ( f, e1 en + f, en e1 ), n

|An | f =

1 ( f, e1 e1 + f, en en ). n

so that

It is easily checked that the An ’s are summable in the weak operator topology, but that the |An |’s are not, since |An | e1 , e1 is the harmonic series. Thus, setting F({xn }) = |An | will not deﬁne an operator-valued measure. However, setting Bn f =

1 f, e1 e1 + f, en en n2

yields a sequence of positive operators that are summable in the weak operator topology and satisfy Bn ± An ≥ 0. Thus, setting F({xn }) = Bn yields a positive operator-valued measure such that F ± E are positive. It is easy to see that the linear map φ: C(X ) → B(H) induced by E is completely bounded, but not of ﬁnite total variation. Note that the decomposition of An = (Bn + An )/2 − (Bn − An )/2 into the dif− ference of two positive matrices is not the usual decomposition, An = A+ n − An , + − for if it were, then necessarily, we should have Bn = An + An = |An |. − In fact, since {|An |} is not summable, either {A+ n } or {An } must not be summable. This happens in spite of the fact that we can decompose each An into a difference of positive operators, namely (Bn ± An )/2, with both of these sequences of positive operators summable.

Bimodule Maps At the moment we have many results about completely bounded maps, but few tools for determining when bounded maps are completely bounded. For bimodule maps there is a nice theory. Let C ⊆ B be C ∗ -algebras. We shall call C matrically norming for B provided for every (Bi j ) in Mn (B) and every n we have that n (Bi j ) = sup Ci Bi j D j : Ci , D j ∈ C i, j=1

Bimodule Maps 107 n Ci Ci∗ ≤ 1 and nj=1 D ∗j D j ≤ 1. This condition looks strange with i=1 until one realizes that   D1  ..  Ci Bi j D j = (C1 , . . . , Cn )(Bi j )  .  i, j Dn and the above inequalities are simply the requirement that the row of C’s and column of D’s each have norm less than or equal to 1. Thus   D1  .  Ci Bi j D j ≤ (C1 , . . . , Cn )(Bi j )  ..  ≤ (Bi j ), i, j Dn and C is matrically norming for B precisely when we can achieve (Bi j ) as the supremum of such products. Note that C plays the same role as the complex numbers play for computing the norms of scalar matrices. For an easy example of this phenomenon, consider B = C(X ), and let C be the constant functions, which can be identiﬁed with C. By compactness we have that for ( f i j ) ∈ Mn (C(X )) there exists x0 with ( f i j ) = ( f i j (x0 )). One can then choose unit vectors h = (c1 , . . . , cn ), k = (d1 , . . . , dn ) such that ( f i j (x0 )) = | i, j ci f i j (x0 ) d j | and hence ( f i j ) = i, j ci f i j d j . Proposition 8.6. Let A, B, and C be C ∗ -algebras with C a subalgebra of both A and B. If C is matrically norming for B and φ: A → B is a C-bimodule map, then φ = φcb . Proof. For any (Ai j ) in Mn (A) we have that (φ(Ai j )) = sup{ Ci φ(Ai j )D j : Ci , D j ∈ C, Ci Ci∗ ≤ 1 and D ∗j D j ≤ 1}. But Ci Ai j D j ≤ φ · (Ai j ) Ci φ(Ai j )D j = φ and the result follows.

Since every linear map is a C-bimodule map, the above result gives another proof of the fact that for maps into B = C(X ) the norm and cb norm are the same. There is another instance in the theory of completely bounded maps where the usual decomposition of a self-adjoint operator A into A+ and A− is not the

108

Chapter 8. Completely Bounded Maps

“best” decomposition into a difference of positive operators. This occurs in the theory of Schur products.

Schur Products Revisited Let A be in Mn , and let S A : Mn → Mn be given by S A (B) = A ∗ B. Note that if D = diag(di ) is a diagonal matrix, then S A (D B) = (ai j di bi j ) = DS A (B) and similarly S A (B D) = S A (B)D. Thus, Schur product maps are bimodule maps over the C ∗ -algebra Dn of diagonal matrices. In fact, by Exercise 4.4, if φ: Mn → Mn is a Dn -bimodule map, then φ = S A for some matrix A. Using the representation theorem for completely bounded maps and Proposition 8.6, we can now characterize the norm of Schur product maps. Theorem 8.7. Let A = (ai j ) be in Mn . Then the following are equivalent: (i) S A ≤ 1, (ii) S A cb ≤ 1, (iii) there exist 2n vectors x1 , . . . , xn , y1 , . . . , yn , in the unit ball of some Hilbert space such that ai j = x j , yi . Proof. Clearly (ii) implies (i). To see that (i) implies (ii), we will prove that Dn is matrically norming for Mn . Suppose that we can prove that for every h 1 , . . . , h k in Cn with h 1 2 + · · · + h k 2 ≤ 1 there exists a unit vector h and D1 , . . . , Dk such that h j = D j h for all j and D ∗j D j ≤ 1. Then given B = (Bi j ) in Mk (Mn ), there will exist h 1 , . . . , h k , h˜ 1 , . . . , h˜ k with h 1 2 + · · · + h k 2 = h˜ 1 2 + · · · + h˜ k 2 = 1 such that B = Bi j h˜ j , h i . i, j

For these vectors we choose D1 , . . . , Dk , D˜ 1 , . . . , D˜ k , h, and h˜ as above and ∗ ˜ Di h. Consequently, B = ˜ ﬁnd that B = i, j Bi j D˜ j h, i, j Di Bi j D j , and we will have that Dn is matrically norming for Mn . To establish the claim, let h 1 , . . . , h k be as above, and let e denote the vector all of whose entries are 1’s. Set H j = diag(h j ), so that h j = H j e and ∗ H j H j ; set h = P 1/2 e and D j = H j P −1/2 . Tr(H j∗ H j ) = h j 2 . Let P = ∗ ∗ −1/2 ( H j H j )P −1/2 = I, D j h = (H j P −1/2 )(P 1/2 e) = Then Dj Dj = P 2 h j , and h2 = P 1/2 e, P 1/2 e = Pe, e = Tr(P) = j h j ≤ 1, and we have shown the claim. Thus, (i) implies (ii), and S A = S A cb for every Schur product map. To prove that (ii) implies (iii), use Theorem 8.4 to write S A (B) = V1∗ π (B)V2 where Vi : Cn → H, i = 1, 2, are isometries and π : Mn → B(H) is a

Schur Products Revisited

109

∗-homomorphism. Let e1 , . . . , en be the canonical basis for Cn , set E j = E 1 j and deﬁne 2n vectors in H via x j = π(E j )V2 e j ,

j = 1, . . . , n,

yi = π(E i )V1 e1 ,

i = 1, . . . , n.

We have x j 2 = V2∗ π(E ∗j · E j )V2 e j , e j = V2∗ π(E j, j )V2 e j , e j ≤ 1,

j = 1, . . . , n,

and similarly that yi 2 ≤ 1,

i = 1, . . . , n.

Let ai, j denote the (i, j)th entry of A; then x j , yi = V1∗ π(E i∗ E j )V2 e j , ei = V1∗ π(E i, j )V2 e j , ei = S A (E i, j )e j , ei = ai, j . Thus, if S A cb ≤ 1, then there is a Hilbert space and 2n vectors in its unit ball such that A = (x j , yi ). Thus, (ii) implies (iii). Finally, we prove that (iii) implies (i). Assume that A = (x j , yi ), where x j ≤ 1 and yi ≤ 1, 1 ≤ i, j ≤ n, are vectors in a Hilbert space H. Given B ∈ Mn , h = (α1 , . . . , αn ), and k = (β1 , . . . , βn ) in Cn , we must prove that |S A (B)h, k| ≤ Bhk. We have ¯ |S A (B)h, k| = x j , yi H bi, j α j β i i, j     α1 x 1 β1 y1 = (bi, j IH )  ... ,  ...  αn x n βn yn Hn 1/2 1/2 ≤ (bi j IH )· βi yi 2 α j x j 2 ≤ B · h · k, since α j x j ≤ |α j | and βi yi ≤ |βi | for all i and j. Thus, S A ≤ 1, and the proof of the theorem is complete.

110

Chapter 8. Completely Bounded Maps

We are now in a position to characterize completely bounded Schur product maps on B(2 ). We leave the analogous result for completely positive maps to the exercises. Corollary 8.8. Let A = (ai j )i,∞j=1 be an inﬁnite matrix. Then the following are equivalent: (i) S A : B(2 ) → B(2 ) is bounded with S A ≤ 1, (ii) S A : B(2 ) → B(2 ) is completely bounded with S A cb ≤ 1, ∞ (iii) there exists a Hilbert space H and vectors {x j }∞ j=1 and {yi }i=1 in H with x j ≤ 1 and yi ≤ 1 for all i and j, such that ai j = x j , yi . Proof. Given B = (bi j )i,∞j=1 , let Bn = (bi, j )i,n j=1 . It is easily checked that B is a bounded operator on 2 if and only if supn Bn is ﬁnite and that in this case B = supn Bn . Consequently, S A is bounded if and only if supn S An is ﬁnite and S A = supn S An . Similarly, one can see that S A cb = supn S An cb and so S A = S A cb , by applying Theorem 8.7. The proofs that (ii) implies (iii) and (iii) implies (i) are as in Theorem 8.7. From the above two results we see that for Schur product maps, S A = S A cb = inf sup{x j yi }, X,Y i, j

where the inﬁmum is over all sets of vectors X = {x j }, Y = {yi } satisfying x j , yi = ai j . Moreover, the inﬁmum is attained. The following criterion for an inﬁnite matrix to give rise to a bounded Schur product map is often useful. Theorem 8.9. Let A = (ai j )i,∞j=1 , and set bi j = ai j − ai+1, j − ai, j+1 + ai+1, j+1 . If the limits lim j ai j = si , limi ai j = t j , limi si , lim j t j all exist and {bi j } is absolutely summable, then S A : B(2 ) → B(2 ) is completely bounded. Proof. First note that ∞ ∞ k= j =i

b,k =

∞

ai,k − tk − ai,k+1 + tk+1

k= j

= ai, j − si − t j + lim tk , k

Schur Products Revisited

111

while ∞ ∞

b,k = ai j − si − t j + lim s .

=i k= j

Hence the absolute summability of {bi j } implies that limi si = lim j t j = w. Choose complex numbers {ci j }, {di j } such that |ci j | = |di j | and ci j d¯i j = bi j . Fix a Hilbert space and a set of orthonormal vectors {ei j }i,∞j=1 ∪ {e, f, g}. Deﬁne x j = −we + t j f + g +

∞ ∞

ck ek ,

k= j =1

yi = e + f + si g +

∞ ∞

dk ek .

k=1 =i

Then it is easily checked that {x j } and {yi } are norm-bounded sets with x j , yi = −w + t j + si +

∞ ∞

bk = ai j ,

k= j =i

and the proof is complete.

It is interesting to see how the decomposition in Theorem 8.3 relates to Schur product maps. If A = (x j , yi ) with x j ≤ 1, yi ≤ 1 for all i, j, then let P1 = (x j , xi ), P2 = (y j , yi ), and note that P1 A ≥ 0. A P2 By Exercise 8.7, we have that S P1 and S P2 are completely positive and this is the decomposition of Theorem 8.3 for S A . It is interesting to see how Wittstock’s decomposition of S A for A self-adjoint relates to the usual decomposition of a self-adjoint matrix as a difference of positive matrices. Suppose that A = A∗ , so that the usual decomposition of A into a difference of positive matrices is given by A = 12 (|A| + A) − 12 (|A| − A) = A+ − A− . From this it follows that S|A| ± S A ≥ 0, that S A = S A+ − S A− is a decomposition of S A as the difference of two completely positive maps and that S A cb ≤ S|A| cb . If equality holds in this last inequality, then S|A| could play the role of ψ in the Wittstock decomposition theorem for φ = S A . However, we shall see [Exercise 8.7(vii)] that, in general, S A cb < S|A| cb . Thus, the usual decomposition of a self-adjoint matrix into a difference of positive matrices, while minimal in some senses, is not minimal for the Wittstock decomposition of S A . This makes it somewhat surprising that when A = A∗ ,

112

Chapter 8. Completely Bounded Maps

then there always is a positive matrix P with P ± A ≥ 0 and S P cb = S A cb [Exercise 8.7(v)]. Let M be an operator space, let φ: M → B, and let and SM be as in Lemma 8.1. It is easy to see that if φk ≤ 1, then the proof of Lemma 8.1 shows that is k-positive. Thus, if φ: M → Mn and φ2n ≤ 1, then : SM → M2 (Mn ) = M2n will be 2n-positive, and consequently completely positive. But, if is completely positive, then φ must be completely contractive. Hence, φ2n = φcb for maps into Mn . Because of the special nature of the subspace SM , it turns out that if is n-positive, then it is completely positive, and consequently, φn = φcb . To obtain this more delicate result, we need a preliminary lemma. Lemma 8.10. Let A be a C ∗ -algebra, and let P = (Pi, j )i,2nj=1 be a positive element in M2n (M2 (A)) where ai, j bi, j Pi, j = , i, j = 1, . . . , 2n. ci, j di, j Then

ai, j

bi, j+n

ci+n, j

di+n, j+n

n i, j=1

is positive in Mn (M2 (A)). Proof. Set A = (ai, j ), B = (bi, j ), C = (ci, j ), and D = (di, j ) for i, j = 1, . . . , 2n. After the canonical shufﬂe, P becomes [ CA DB ], and so this latter matrix is positive. Now partition each of the 2n × 2n matrices into 2 × 2 matrices consisting of n × n blocks in the natural fashion. That is, A11 A12 , A= A21 A22 where A11 = (ai, j )i,n j=1 , A12 = (ai, j+n )i,n j=1 , A21 = (ai+n, j )i,n j=1 , and A22 = (Ai+n, j+n )i,n j=1 with the same deﬁnitions for B, C, and D. Thus, 

A C

is positive.

A11 A B  21 = D C11 C21

A12 A22 C12 C22

B11 B21 D11 D21

 B12 B22    D12  B22

Schur Products Revisited

113

A moment’s reﬂection shows that if such a matrix is positive, then the four corners will form a positive matrix, that is, A11 B12 C21 D22 will be positive. The result follows by performing the canonical shufﬂe on this last matrix. Proposition 8.11 (Smith). Let M be an operator space, and let φ: M → Mn be a linear map. Then φn = φcb . Proof. It is sufﬁcient to assume that φn = 1 and prove that : SM → M2n is completely positive, where and SM are deﬁned as in Lemma 8.1. By Theorem 5.1, to prove is completely positive, it is sufﬁcient to prove that the associated linear functional S : M2n (SM ) → C is positive. Thus, let P = (Pi, j )i,2nj=1 be positive in M2n (SM ), with λi, j ai, j Pi, j = ∗ , bi, j µi, j and calculate 2n · S (P) =

n

(Pi, j )(i, j) +

i, j=1

+

n

n

n

+

n

φ(ai, j+n )(i, j+n)

i, j=1

n

φ(bi, j )∗(i+n, j) +

i, j=1

=

n

(Pi+n, j+n )(i+n, j+n)

i, j=1

λi,i +

i=1

n

(Pi+n, j )(i+n, j) +

i, j=1

=

(Pi, j+n )(i, j+n)

i, j=1

n

µi+n,i+n

i=1

φ(Ai, j )(i, j) ,

i, j=1

where

Ai, j =

λi, j

ai, j+n

∗ bi+n, j µi+n, j+n

.

By the above lemma, A = (Ai, j ) is a positive element of Mn (SM ), and if we set x = e1 ⊕ · · · ⊕ en , then the last sum above is easily recognized as n (A)x, x.

114

Chapter 8. Completely Bounded Maps

But this last expression is positive, because the assumption that φn = 1 is enough to guarantee that is n-positive, as observed in the remarks proceeding Lemma 8.10. The counterpart of Theorem 3.14 for completely bounded maps is not true. If φ: Mn → B, then, in general, φn = φcb , and in fact, there does not exist any integer m = m(n) such that φm = φcb for all φ. This follows from work of Haagerup [108]. Combining Proposition 8.11 with Exercise 3.10(ii), we see that for maps into Mn , φcb ≤ nφ, which gives another proof of Exercise 3.11.

Notes Wittstock obtained his decomposition theorem for completely bounded maps in [242] and proved the extension theorem for completely bounded maps in [243]. His proofs used his theory of matricial sublinear functionals, which is a generalization of the Hahn–Banach theorem to set-valued mappings into B(H). Haagerup [107] obtained these same results by exploiting the correspondence between maps from M into Mn and linear functionals on Mn (M), as was done for completely positive maps in Chapter 6. The difﬁcult part of this approach in the completely bounded case is that the norm of the associated linear functional and the cb norm of the mapping into Mn are not very closely related. Thus, for example, to obtain an extension of a map from M into Mn to A into Mn , where M ⊆ A, with the same cb norm, one must extend the associated linear functional on Mn (M) to Mn (A), but in a very particular fashion. This technical stumbling block was the main obstruction to the theory of completely bounded maps developing simultaneously with the theory of completely positive maps. To overcome this difﬁculty, Haagerup used a technique that eventually led to the deﬁnition of the Haagerup tensor norm, which is studied in Chapter 17. The proofs that we have presented here using the off-diagonal technique appeared in [158] and [159]. The generalization of Stinespring’s representation theorem to completely bounded maps is referred to as the factorization theorem for completely bounded maps in some other texts. This result ﬁrst appeared in [107] and [159]. Although this result can be deduced from Wittstock’s decomposition and extension theorem, it is a bit tricky to get the statement about the equality of the norms via that approach. Even in the abelian case and with C for the range, it is difﬁcult to express the norm of a measure in terms of inequalities involving its real and

Exercises

115

imaginary parts. Arguably the only way to achieve this in the noncommutative case is via the off-diagonal trick. The results in [107], [159], and [243] were obtained independently and within about a six-month time period. It is certain that Wittstock obtained his proofs ﬁrst, and in any case his early work was the inspiration for [107] and [159]. Proposition 8.11 was obtained by Smith [214]. Haagerup [107] was the ﬁrst to prove that for Schur product maps, S A = S A cb . Smith [217] later gave a simpler proof that contained the idea of a matricially norming subalgebra. This concept has subsequently been developed further by Pop, Sinclair, and Smith [195] and has many applications to computing cohomology. Theorem 8.9 is a slight variation of a result of Bennett [14]. The generalization of the off-diagonalization technique to C-bimodule maps (Exercise 8.6) is in Suen [225]. Wittstock had proved the decomposition and extension theorems in this more general setting earlier ([242] and [243]). Exercise 8.8 is proven in Haagerup [108], where a partial converse of the Wittstock decomposition theorem is obtained. Namely, if a von Neumann algebra has the property that the span of the completely positive maps from it into itself is the completely bounded maps, then it is injective. An example is given by Huruya [125] of a C ∗ -algebra that is not injective, but is such that every completely bounded map into it is in the span of the completely positive maps. However, this decomposition does not meet the norm inequality in the statement of the Wittstock decomposition property. It is known that a C ∗ -algebra that has the Wittstock decomposition property, including the norm inequality, is injective. See [220] and [193] for more on this topic. Smith [214] proved that the C ∗ -algebra C([0, 1]) has the property that not every completely bounded map from it to itself is in the span of the completely positive maps.

Exercises 8.1 Show that Re(φn ) = (Re φ)n , and that (φ ∗ )n = (φn )∗ . 8.2 Let φ: M → B, let H, K be in Mn , and let A be in Mn (M). Prove that φn (H · A · K ) = H · φn (A) · K . Thus, φn : Mn (M) → Mn (B) is an Mn bimodule map. 8.3 Verify the claim of Theorem 8.4. 8.4 Show that if φ is completely bounded, and φ(a) = V1∗ π (a)V2 is the representation of Theorem 8.4 with V1 = V2 , then setting φi (a) = Vi∗ π (a)Vi yields the map of Theorem 8.3.

116

Chapter 8. Completely Bounded Maps

8.5 Prove that the conclusions of Theorems 8.2, 8.3, and 8.5 still hold when the range is changed from B(H) to an arbitrary injective C ∗ -algebra. 8.6 Let A, B, and C be C ∗ -algebras with unit, with C contained in both A and B, and 1A = 1C , 1B = 1C . Let M ⊆ A be a subspace such that c1 · M · c2 ⊆ M for all c1 , c2 in C, and set c1 a S= : a, b ∈ M, c1 , c2 ∈ C . b∗ c2 (i) Prove that if φ: M → B is a completely contractive C-bimodule map, then : S → M2 (B) deﬁned by c1 c1 a φ(a) = b∗ c2 φ(b)∗ c2 is completely positive. (ii) Prove that if B is injective, then the conclusions of Theorems 8.2, 8.3, and 8.5 still hold with the additional requirement that the maps be C-bimodule maps. (Hint: Use Theorem 3.18.) 8.7 Let A = (ai j )i,∞j=1 . Prove that the following are equivalent: (i) S A : B(2 ) → B(2 ) is positive, (ii) S A : B(2 ) → B(2 ) is completely positive, (iii) there exists a Hilbert space H and a bounded sequence of vectors {xi } in H such that ai j = x j , xi . 8.8 Let S A : Mn → Mn be the Schur product map S A (B) = A ∗ B. Let A = H + i K be the decomposition of A into its real and imaginary parts. (i) Show that Re(S A ) = S H . (ii) Prove that there exist positive matrices P1 and P2 such that φi = S Pi , i = 1, 2, satisfy the conclusions of Theorem 8.3 for φ = S A . (Hint: Recall Exercise 4.4.) (iii) Prove that P1 and P2 are such a pair of positive matrices if and only if S P1 cb = S P2 cb = S A cb and [ AP1∗ PA2 ] is positive. (iv) Let S A cb ≤ 1, and write A = (x j , yi ) as in Theorem 8.7. Show that P1 = (y j , yi ), P2 = (x j , xi ) are such a pair of positive matrices. (v) Let d(B) denote the maximum diagonal element of a matrix. Conclude that P1 A P1 A : ≥0 . S A cb = inf d A∗ P2 A∗ P2

Exercises

117

(vi) Show that

A

A∗

|A|

|A∗ |

≥ 0.

(vii) Give an example where S A cb < max{d(|A∗ |), d(|A|)}, with A = A∗ . 8.9 (Haagerup) Let φ1 , φ2 , φ: A → B. Prove that : M2 (A) → M2 (B), deﬁned by φ1 (a) φ(b) a b = , c d φ ∗ (c) φ2 (d) is completely positive if and only if : A → M2 (B), deﬁned by φ1 (a) φ(a) (a) = , φ ∗ (a) φ2 (a) is completely positive. 8.10 Let : M2 (Mn ) → M2 (B) be given by + φ (a) φ(b) a b , = c d φ ∗ (a) φ − (d) and let E = (E i, j ) be in Mn (Mn ), where E i, j are the canonical matrix units. Prove that is completely positive if and only if φn+ (E) φ(E) φn∗ (E) φn− (E) is positive. 8.11 Let φ: Mn → B, set B = (φ(E i, j )) in Mn (B), and let |B ∗ | = ( pi, j ), |B| = (qi, j ). Deﬁne linear maps φ1 , φ2 : Mn → B by φ1 (E i, j ) = pi, j , φ2 (E i, j ) = qi, j . (i) Prove that : M2 (Mn ) → M2 (B), deﬁned by φ1 (a) φ(b) a b , = c d φ ∗ (c) φ2 (d) is completely positive. (ii) Prove that φ2cb ≤ p11 + · · · + pnn 2 · q11 + · · · + qnn 2 , and that this estimate is sharp.

118

8.12 8.13

8.14

8.15

8.16

Chapter 8. Completely Bounded Maps (iii) Let B ∗ B = (ci, j ). Prove that c11 + · · · + cnn ≤ n · φ2cb , and that this estimate is sharp. Prove directly that if φ: C(X ) → B(H) has ﬁnite total variation, then φ is completely bounded. Let E be a completely bounded, operator-valued measure. Prove that E has ﬁnite total 2-variation, i.e., that sup{ |E(Bi )|2 : Bi disjoint, Borel} < +∞. (Sakai) Let A be a C ∗ -algebra, and let f : A → C be a bounded, selfadjoint linear functional with f ≤ 1. Show that f = f 1 − f 2 , with f 1 and f 2 positive linear functionals and f 1 + f 2 = 1. Let A = (ai j )i,∞j=1 , and assume that S A is bounded. Prove that if limi lim j ai j and lim j limi ai j both exist, then they must be equal. (Hint: Write ai j = x j , yi , and use the weak∗ compactness of the ball in Hilbert space.) Prove that if for every B = (bi j ) on B(2 ) the matrix 0, i > j, where ci j = C = (ci j ) bi j , i ≤ j, is in B(2 ), then S A is bounded, where A = (ai j ),

ai j =

0, 1,

i> j . i≤ j

Use Exercise 8.15 to show that S A is not bounded, and deduce that there exists B in B(2 ) for which C is unbounded. 8.17 Fix a subset P ⊆ {1, . . . , n} × {1, . . . , n}, and deﬁne a subspace M P ⊆ / P. Mn via B ∈ M P if and only if B = (bi j ) with bi j = 0 whenever (i, j) ∈ Show that M ⊆ Mn is a Dn -bimodule if and only if M = M P for some P. 8.18 Let P, M P be as in Exercise 8.17, and assume we are given complex numbers (ai j ) for every (i, j) ∈ P. Such a matrix is called partially deﬁned. Show that there is a well-deﬁned Dn -bimodule map φ: M p → M p given by φ((bi j )) = (ai j bi j ) and that every Dn -bimodule map arises this way. Prove that φ = φcb = inf X,Y sup(i, j)∈P {x j yi }, where the inﬁmum is over every set X = {x j }, Y = {yi } satisfying x j , yi = ai j for all (i, j) ∈ P. Given a partially deﬁned matrix A = (ai j ) as above, a choice of the remaining undeﬁned entries is called a completion of A. 8.19 Show that M P is an operator system if and only if (i, i) ∈ P for 1 ≤ i ≤ n and if (i, j) ∈ P then ( j, i) ∈ P. Let M P be an operator system, and let φ: M P → M P and A = (ai j ) be as in Exercise 8.18. Prove that the

Exercises

119

following are equivalent: (i) φ is positive, (ii) φ is completely positive, (iii) A has a positive completion. 8.20 Say E has ﬁnite total p-variation if sup{ |E(Bi )| p : Bi disjoint, Borel} < +∞. Does E have ﬁnite total 1-variation if and only if its associated map, φ: C(X ) → B(H), has ﬁnite total variation? If E is completely bounded, must E have ﬁnite total p-variation for all p > 1?

Chapter 9 Completely Bounded Homomorphisms

In Chapter 7 we saw that if B is a C ∗ -algebra with unit, and A is a subalgebra of B containing 1B , then the unital, completely contractive homomorphisms of A into B(H) are precisely the homomorphisms with B-dilations. In this chapter, we prove that the unital homomorphisms of A, which are similar to homomorphisms with B-dilations, are precisely the completely bounded homomorphisms. If A is a C ∗ -algebra, then every unital, contractive homomorphism is a positive map and hence a ∗-homomorphism. Thus, for unital maps of C ∗ -algebras, the sets of contractive homomorphisms, completely contractive homomorphisms, and ∗-homomorphisms coincide. In this case, the above result says that a unital homomorphism of a C ∗ -algebra is similar to a ∗-homomorphism if and only if it is a completely bounded homomorphism. Let’s begin with a simple observation. Suppose that S is a similarity, ρ is a homomorphism of some operator algebra A, and π (·) = S −1 ρ(·)S is a completely contractive homomorphism. Letting Sn denote the direct sum of n copies of S, we have that Sn−1 = (S −1 )n , Sn = S, and ρn (·) = Sn πn (·)Sn−1 . Thus, ρ is completely bounded with ρcb ≤ S −1 · S. So any homomorphism that is similar to a completely contractive homomorphism is necessarily completely bounded with ρcb ≤ inf{S −1 · S: S −1 ρ(·)S is completely contractive}. The next result proves that not only is the converse of the above statement true, but the above inﬁmum is achieved and gives the cb norm of ρ. Theorem 9.1. Let A be an operator algebra, and let ρ: A → B(H) be a unital, completely bounded homomorphism. Then there exists an invertible operator S with S · S −1 = ρcb such that S −1 ρ(·)S is a completely contractive 120

Chapter 9. Completely Bounded Homomorphisms

121

homomorphism. Moreover, ρcb = inf{R −1 · R: R −1 ρ(·)R is completely contractive}. Proof. The remarks preceding the statement of the theorem yield the last equality. Let A be contained in a C ∗ -algebra B. By the extension and representation theorems for completely bounded maps, we know that there exists a Hilbert space K, a ∗-homomorphism π: B → B(K), and two bounded operators Vi : H → K, i = 1, 2, with ρcb = V1 · V2 , such that ρ(a) = V1∗ π (a)V2 for all a in A. For h in H, we deﬁne ρ(ai )h i = h, ai ∈ A, h i ∈ H , |h| = inf π(ai )V2 h i : where the inﬁmum is taken over all ﬁnite sums. It is easy to see that |·| deﬁnes a seminorm on H. If h = ρ(ai )h i , then h = ρ(ai )h i = V1∗ π(ai )V2 h i ≤ V1∗ · π(ai )V2 h i , and thus h ≤ V1∗ · |h|. Similarly, the equation ρ(1)h = h yields |h| ≤ V2 · h. Thus, |·| is equivalent to the original norm on H. Hence |·| is also a norm on H, and (H, |·|) is complete. We claim that (H, |·|) is also a Hilbert space. By the Jordan–von Neumann theorem, it is enough to verify that |·| satisﬁes the parallelogram law |h + k|2 + |h − k|2 = 2(|h|2 + |k|2 ) for all h and k in H. Let h = ρ(ai )h i , k = ρ(bi )ki , then h ± k = ρ(ai )h i ± ρ(bi )ki , and so 2 π(ai )V2 h i + π(bi )V2 ki |h + k|2 + |h − k|2 ≤ 2 + π(ai )V2 h i − π(bi )V2 ki 2 2 π(ai )V2 h i + 2 π (bi )V2 ki . = 2 Taking the inﬁmum over all such sums yields |h + k|2 + |h − k|2 ≤ 2|h|2 + 2|k|2 . The other inequality follows by substituting h + k and h − k for h and k, respectively. Hence, (H, |·|) is also a Hilbert space.

122

Chapter 9. Completely Bounded Homomorphisms

Let S: (H, |·|) → (H, · ) be the identity map. Then S is bounded and invertible, S −1 · S ≤ V1∗ · V2 = ρcb , and S −1 ρ(·)S is just the homomorphism ρ, but with respect to the |·|-norm. Thus, to complete the proof of the theorem, it is sufﬁcient to prove that ρ is completely contractive with respect to this new norm. (To see this, let U : (H, · ) → (H, | · |) be a unitary, and set R = U S.) It is not difﬁcult to see that ρ(·) is contractive with respect to the |·|-norm, for if a ∈ A, h ∈ H, and h = ρ(ai )h i , then π(aai )V2 h i ≤ a · π (ai )V2 h i , |ρ(a)h| = ρ(aai )h i ≤ and so |ρ(a)h| ≤ a · |h|. To see that ρ(·) is completely contractive in the |·|-norm, ﬁx an integer n, and let Hˆ = H ⊕ · · · ⊕ H (n copies). Let |·|n denote the Hilbert space norm on Hˆ

induced by |·|, that is,

ˆ 2n = |h 1 |2 + · · · + |h n |2 |h| ˆ We must prove that if A = (ai, j ) is in Mn (A), then for hˆ = (h 1 , . . . , h n ) in H. ˆ n ≤ A · |hˆ n |. |ρn (A)h| Consider Hˆ with its old norm, h2n = h 1 2 + · · · + h n 2 . ˆ · n ) is a completely Since ρ is completely bounded, ρn : Mn (A) → B(H, bounded homomorphism, and so, by the ﬁrst part of the proof, we can endow Hˆ with yet another norm |·| such that ρn is contractive in the |·| -norm. By the ﬁrst part of the proof, to deﬁne |·| , we need a Stinespring representation of ρn . To this ˆ i = 1, 2, end, let Kˆ = K ⊕ · · · ⊕ K (n copies), Vˆ i = Vi ⊕ · · · ⊕ Vi : Hˆ → K, ˆ so that ρn (·) = Vˆ ∗ πn (·)Vˆ 2 . If we deﬁne and πn : Mn (A) → B(K), 1 ˆ = inf ρn (Ai )Vˆ 2 hˆ i : ρn (Ai )hˆ i = hˆ , |h| n

where the inﬁmum is taken over all ﬁnite sums, then ρn (·) will be contractive in |·| . The proof of the theorem is now completed by showing that |·| = |·|n , which we leave to the reader (Exercise 7.1). Corollary 9.2. (Haagerup). Let A be a C ∗ -algebra with unit, and let ρ: A → B(H) be a bounded, unital homomorphism. Then ρ is similar to a ∗-homomorphism if and only if ρ is completely bounded. Moreover, if ρ is completely bounded, then there exists a similarity S with S −1 ρ(·)S a ∗homomorphism and S −1 · S = ρcb .

Chapter 9. Completely Bounded Homomorphisms

123

The above result sheds considerable light on a conjecture of Kadison [128]. Kadison’s conjecture is that every bounded homomorphism of a C ∗ -algebra into B(H) is similar to a ∗-homomorphism. Haagerup’s result shows that Kadison’s conjecture is equivalent to determining whether or not bounded homomorphisms are necessarily completely bounded. At the present time the similarity question is still open, although there are a number of deep partial results (see [12], [35], [55], and [106]). In Chapter 19 we will return to this question and study the recent contributions of Pisier to this problem. There are several important cases where bounded homomorphisms of a C ∗ algebra are completely bounded. These are closely tied to the theory of group representations. Let G be a locally compact, topological group, and let L ∞ (G) denote the C ∗ -algebra of bounded, measurable functions on G. Given g in G, we deﬁne the (right) translation operators Rg : L ∞ (G) → L ∞ (G) by (Rg f )(g ) = f (g g). A state m: L ∞ (G)∗ → C is called a (right) invariant mean on G if m(Rg f ) = m( f ) for all g in G. The group G is called amenable if there exists an invariant mean on G. Note that if G is compact, then it is amenable, since we may deﬁne f (g) dg, m( f ) = G

where dg denotes Haar measure. Another important class of groups that are amenable are the commutative groups. To see this, note that the adjoint maps, Rg∗ : L ∞ (G)∗ → L ∞ (G)∗ are weak∗ -continuous and map states to states. Since the space of states is weak∗ compact and convex, and since the maps Rg∗ form a commutative group of continuous maps on this space, by the Markov–Kakutani ﬁxed point theorem [78] there will exist a ﬁxed point. This ﬁxed point is easily seen to be an invariant mean. The importance of the existence of an invariant mean is best illustrated in the following result. Theorem 9.3 (Dixmier). Let G be an amenable group, and let ρ: G → B(H) be a strongly continuous homomorphism with ρ(e) = 1, such that ρ = sup{ρ(g): g ∈ G} is ﬁnite. Then there exists an invertible S in B(H) with S · S −1 ≤ ρ2 such that S −1 ρ(g)S is a unitary representation of G. Proof. Let m denote an invariant mean on G, and note that for each pair of vectors x, y in H, the function f x,y (g) = ρ(g)x, ρ(g)y is a bounded, continuous function on G. Set x, y1 = m( f x,y ), and note that since the map (x, y) → f x,y is sesquilinear, , 1 deﬁnes a sesquilinear form on H. Also, since m is a positive

124

Chapter 9. Completely Bounded Homomorphisms

linear functional, and f x,x is a positive function, this new sesquilinear form is a semideﬁnite inner product on H. We shall show that it is bounded above and below by the original inner product. To see this, set M = ρ and note that since ρ(g) ≤ M and ρ(g)−1 ≤ M, we have that 1/M 2 ≤ ρ(g)∗ ρ(g) ≤ M 2 . Hence, (1/M 2 )x, x ≤ f x,x (g) ≤ M 2 x, x, and applying m to this inequality yields 1 x, x ≤ x, x1 ≤ M 2 x, x. M2 Thus, x, y1 is an equivalent inner product on H. If we let (H, |·|1 ) denote our Hilbert space with the norm induced by this new inner product, then just as in the proof of Theorem 9.1, the map S: (H, · ) → (H, |·|1 ) is bounded and invertible, with S · S −1 ≤ M 2 = ρ2 . Finally, note that with respect to this new inner product, ρ(g)x, ρ(g)y1 = m(Rg f x,y ) = m( f x,y ) = x, y1 , and so ρ(g) is a unitary on (H, |·|1 ).

Corollary 9.4. (Sz.-Nagy). Let T be an invertible operator on a Hilbert space such that T n ≤ M for all integers n. Then there exists an invertible operator S, with S −1 · S ≤ M 2 , such that S −1 T S is a unitary operator. Proof. Let Z denote the group of integers, deﬁne ρ(n) = T n , and apply Theorem 9.3. In a number of cases, similarity results for homomorphisms of C ∗ -algebras can be obtained from Theorem 9.3 by restricting attention to the group of unitaries in the C ∗ -algebra. The following result demonstrates this principle. Lemma 9.5. Let A be a C ∗ -algebra. Then every element in A is a linear combination of at most four unitaries. Proof.√ Let h = h ∗ be a self-adjoint element of A, with h ≤ 1. Then u = h + i 1 − h 2 is easily seen to be unitary, and h = u + u ∗ . This shows that every self-adjoint element is in the span of two unitaries. Using the Cartesian decomposition, we obtain the result.

Chapter 9. Completely Bounded Homomorphisms

125

Lemma 9.6. Let A and B be C ∗ -algebras, and let ρ: A → B be a homomorphism with ρ(1) = 1. If ρ maps unitaries to unitaries, then ρ is a ∗-homomorphism. Proof. Let u be unitary in A; then ρ(u ∗ ) = ρ(u)−1 = ρ(u −1 ) = ρ(u)∗ , and so ρ is self-adjoint on the unitary elements of A. By Lemma 9.5, ρ is self-adjoint on A. Theorem 9.7. Let A be a commutative, unital C ∗ -algebra. If ρ: A → B(H) is a bounded homomorphism, then ρ is completely bounded and ρcb ≤ ρ2 . Proof. Let G denote the commutative group of unitary elements of A. By Theorem 9.3, there is a similarity S, with S · S −1 ≤ ρ2 , such that S −1 ρ(u)S is unitary for all u in G. By Lemma 9.6, S −1 ρ(·)S is a ∗-homomorphism. Corollary 9.8. Let T be an operator on a Hilbert space. Then T is similar to a self-adjoint operator if and only if for some interval, [a, b], there is a constant K such that p(T ) ≤ K · sup{| p(t)|: t ∈ [a, b]} for all polynomials p with real coefﬁcients. By Theorem 9.3 and Lemma 9.6, if A is a C ∗ -algebra, G is an amenable group contained in the group of unitaries of A, and B is the C ∗ -subalgebra generated by G, then for any bounded homomorphism ρ: A → B(H) there is an invertible S, with SS −1 ≤ ρ2 such that S −1 ρ(·)S is a ∗-homomorphism on B. An afﬁrmative answer to Kadison’s conjecture for C ∗ -algebras is known to imply an afﬁrmative answer to two other questions concerning derivations and invariant operator ranges. Consequently, these problems have an analogous status, they have afﬁrmative answers if and only if certain bounded maps are completely bounded, but at the present time, both questions are still open. We discuss the derivation question in this chapter. For a discussion of invariant operator ranges and their connections with completely bounded maps, see [158]. Let A be a C ∗ -algebra and π: A → B(H) a unital ∗-homomorphism. A linear map δ: A → B(H) is called a derivation if δ(AB) = π(A)δ(B) + δ(A)π (B). It is known that every derivation is automatically bounded [199]. If X ∈ B(H), then setting δ(A) = π(A)X − X π (A) deﬁnes a derivation, and such a derivation is called inner. The derivation question asks whether every derivation into B(H) is necessarily inner.

126

Chapter 9. Completely Bounded Homomorphisms

Given a derivation δ, if we deﬁne ρ: A → B(H ⊕ H) by π(A) δ(A) ρ(A) = , 0 π (A) then ρ will be a homomorphism of A. In fact, it is not difﬁcult to see that ρ is a homomorphism if and only if δ is a derivation. Proposition 9.9. The derivation δ is inner if and only if ρ is similar to a ∗-homomorphism. Proof. First, suppose that δ is inner, so that δ(A) = π (A)X − X π (A). Setting

1 −X 1 X −1 S= , , S = 0 1 0 1 we have that

π (A) 0 , = 0 π (A)

Sρ(A)S

−1

which is a ∗-homomorphism. Conversely, if S is a similarity such that γ (A) = S −1 ρ(A)S is a ∗-homomorphism, then set X = SS ∗ . We have that ρ(A)X = Sγ (A)S ∗ = (Sγ (A∗ )S ∗ )∗ = (ρ(A∗ )SS ∗ )∗ = Xρ(A∗ )∗ . Writing

X 11 X= ∗ X 12

X 12 , X 22

the above equation becomes ∗ π(A)X 12 + δ(A)X 22 π(A)X 11 + δ(A)X 12

=

∗ π(A)X 12 π (A)X 22 X 11 π(A) + X 12 δ(A∗ )∗ X 12 π (A) ∗ X 12 π(A) + X 22 δ(A∗ )∗

X 22 π (A)

.

But since X is positive and invertible, X 22 must also be positive and invertible. Equating the (1,2) entries of the above operator matrices yields −1 −1 = (X 12 π(A) − π (A)X 12 )X 22 δ(A) = δ(A)X 22 X 22 −1 −1 = X 12 X 22 π(A) − π (A)X 12 X 22 ,

since X 22 commutes with A.

Chapter 9. Completely Bounded Homomorphisms

127

Corollary 9.10 (Christensen). Let A be a unital C ∗ -algebra, π : A → B(H) a unital ∗-homomorphism, and δ: A → B(H) a derivation. Then δ is inner if and only if δ is completely bounded. Proof. By using the canonical shufﬂe, it is easy to see that ρ is completely bounded if and only if δ is completely bounded. But by Proposition 9.9 and Corollary 9.2, δ is inner if and only if ρ is completely bounded. Thus, for a given C ∗ -algebra A, we see that if every bounded homomorphism of A is similar to a ∗-homomorphism, then every derivation of A is inner. Kirchberg [132] has proven the converse of this result. Namely, if a C ∗ -algebra has the property that every derivation into B(H) is inner, then necessarily every bounded homomorphism is similar to a ∗-homomorphism. Thus, for C ∗ -algebras, Kadison’s conjecture and the problem of determining if every derivation from a C ∗ -algebra into the bounded operators on some Hilbert space is inner are equivalent. Turning our attention to some non-self-adjoint algebras, Theorem 9.1 can be used to give a characterization of the operators that are similar to a contraction. Let P(D) denote the algebra of polynomials with the norm it inherits as a subalgebra of C(T). Theorem 9.11. Let T ∈ B(H). Then T is similar to a contraction if and only if the homomorphism ρ: P(D) → B(H) deﬁned by ρ( p) = p(T ) is completely bounded. Moreover, if this is the case, then ρcb = inf{S · S −1 : S −1 T S ≤ 1}, and the inﬁmum is attained. Proof. If ρ is completely bounded, then there is a similarity S, with S −1 · S = ρcb , such that S −1 ρ(·)S is completely contractive. Thus, S −1 T S = S −1 ρ(z)S ≤ z = 1, where z is the coordinate function. Conversely, if R = S −1 T S is a contraction, then θ: P(D) → B(H), given by θ ( p) = p(R), is a contractive homomorphism by von Neumann’s inequality, and is completely contractive by Sz.-Nagy’s dilation theorem. But ρ(·) = Sθ(·)S −1 , and so ρ is completely bounded with ρcb ≤ S · S −1 . The statement about the inﬁmum follows from the corresponding statement in Theorem 9.1.

128

Chapter 9. Completely Bounded Homomorphisms

An operator for which ρ( p) = p(T ) is a bounded homomorphism of P(D), that is, an operator for which the closed unit disk is a K -spectral set, is called a polynomially bounded operator. It is quite natural to refer to the operators for which this ρ is completely bounded as completely polynomially bounded. Halmos [111] conjectured that every polynomially bounded operator is similar to a contraction. The above result shows that Halmos’s conjecture is equivalent to determining whether or not bounded homomorphisms of P(D), or equivalently A(D), are completely bounded. Pisier [191] produced an example, in fact, a whole family of examples, of a polynomially bounded operator that is not completely polynomially bounded. Thus proving that Halmos’s conjecture is not true. We shall present these examples in Chapter 10. Holbrook [122] gave the ﬁrst example of a matrix T , which is similar to a contraction such that ρ < inf{S −1 · S: S −1 T S ≤ 1}, where ρ( p) = p(T ) is the induced homomorphism of P(D). The theory of completely bounded maps at least gives a basis for explaining such phenomena, since the right-hand side of the above equation is ρcb and, in general, there is no reason to expect equality of these two norms. Using the theory of completely bounded maps, there are now quite good estimates on how large the ratio of ρcb to ρ can be when T is an n × n matrix. See for example [193]. The above theory sheds further light on the theory of K -spectral sets. Let X be a compact set in C, let R(X ) be the algebra of quotients of polynomials, and let T be in B(H) with σ (T ) contained in X . Recall that if there is a bounded homomorphism ρ: R(X ) → B(H) deﬁned by ρ(r ) = r (T ) with ρ ≤ K , then X is called a K -spectral set for T . If ρ is completely bounded with ρcb ≤ K , we shall call X a complete K -spectral set for T . Recall also that X is a complete spectral set for T if and only if T has a normal ∂ X -dilation. Translating Theorem 9.1 into this language yields the following result. Corollary 9.12. A set X is a complete K -spectral set for T if and only if T is similar to an operator for which X is a complete spectral set. Moreover, there exists such a similarity satisfying S −1 · S ≤ K . In fact, we have that for ρ as above, ρcb = inf{S −1 · S: X is a complete spectral set for S −1 T S}, and the inﬁmum is attained. The full picture of how these concepts are related is still not complete. As discussed earlier, it is still not known, if a subset X of C is a spectral set for

Chapter 9. Completely Bounded Homomorphisms

129

an operator, whether or not X must be a complete spectral set for the operator. However, we saw in Chapter 5 that higher-dimensional analogues of this question, when say X is a polydisk, can fail. It is known that if a set is a K -spectral set for an operator, then it need not be a complete K -spectral set for the operator, for any K . Halmos’s conjecture was the special case of this question where X = D− , and Pisier’s counterexample shows that the above fails even on this simple domain. However, there exist sets for which it is the case that whenever the set is a K -spectral set for an operator, then it is a complete K -spectral set for some K . For example, when R(X ) is dense in C(∂ X ), then this is the case. But it is unknown if this is the only situation where this occurs. Maps which are deﬁned by some type of analytic functional calculus are often completely bounded. Recall the Riesz functional calculus. If G is an open set containing σ (T ) and ∂G consists of a ﬁnite number of simple, closed, rectiﬁable curves, then for any function in R(G), we have that 1 r (z)(z I − T )−1 dz. r (T ) = 2πi ∂G This functional calculus can be used to derive what is referred to as the generalized Rota model of an operator. Theorem 9.13 (Herrero–Voiculescu). Let T be in B(H), and let G be an open set in the complex plane such that σ (T ) is contained in G and such that ∂G consists of a ﬁnite number of simple, closed, rectiﬁable curves. Then T is similar to an operator that has a normal ∂G-dilation. Moreover, the similarity may be chosen to satisfy 1 −1 (z I − T )−1 |dz|. S · S ≤ 2π ∂G Proof. By Corollary 9.12, we need only prove that G − is a complete K -spectral set for T with 1 (z I − T )−1 |dz|. K = 2π ∂G Let ρ: R(G − ) → B(H) be the homomorphism deﬁned by ρ(r ) = r (T ). We must prove that ρcb ≤ K . By the Riesz functional calculus, 1 r (z) · (z I − T )−1 |dz| ≤ K · r , ρ(r ) ≤ 2π ∂G so ρ ≤ K . Let Tˆ denote the direct sum of n copies of T and note that

130

Chapter 9. Completely Bounded Homomorphisms

(z Iˆ − Tˆ )−1 = (z I − T )−1 . Thus, for (ri, j ) in Mn (R(G − )), we have that 1 ρn ((ri, j )) = (ri, j (z) · (z I − T )−1 ) dz 2πi ∂G 1 = (ri, j (z)) · (z Iˆ − Tˆ )−1 dz, 2πi ∂G and so, by the same inequality as above, ρn ≤ K , which proves that G − is a complete K -spectral set for T . Corollary 9.14. (Rota’s Theorem). Let T be in B(H) with σ (T ) contained in the open unit disk. Then T is similar to a contraction. Moreover, a similarity may be chosen satisfying 1 −1 (z I − T )−1 |dz|. S · S ≤ 2π T Proof. By Corollary 9.12, T is similar to an operator with a unitary dilation, and that operator is necessarily a contraction. The original proof of Rota’s theorem gives much more particular information on the unitary dilation and similarity than this proof does. In fact, the unitary can always be taken to be the bilateral shift, and the invertible S can be chosen to belong to the C ∗ -algebra generated by T (see, for example, [97] or Exercise 9.15). One difference is that the above proof can give quite different estimates on S −1 · S. Recall the spectral radius formula, sup{|z|: z ∈ σ (T )} = lim T n 1/n . n

So if σ (T ) is contained in D, then (eiθ I − T )−1 = e−iθ

n

(e−iθ T )n ,

and the latter series converges by the root test. Thus, we see that the integral estimate on S −1 · S obtained in Corollary 9.14 satisﬁes 1 (z I − T )−1 |dz| ≤ T n . n 2π T The proof of Rota’s theorem in Exercise 9.15 leads to an upper bound on S · S −1 of n T n 2 .

Notes

131

Notes Kadison’s study of homomorphisms of C ∗ -algebras [128] was motivated in part by Dixmier’s study of whether or not bounded representations of groups are similar to unitary representations [72]. Dixmier’s question is known to have a negative answer, that is, there exist bounded representations of groups that are not similar to unitary representations. This result was obtained by Kunze and Stein [133]. The original proof of Sz.-Nagy’s result characterizing operators that are similar to unitaries (Corollary 9.8) contains the elements of Dixmier’s result on amenable groups (Theorem 9.3). This proof is outlined in the exercises (Exercise 9.7) and can be found in [226]. Hadwin [109] proved that a unital homomorphism of a C ∗ -algebra into B(H) is similar to a ∗-homomorphism if and only if the homomorphism is in the span of the completely positive maps from the algebra to B(H), and conjectured that the span of the completely positive maps was the completely bounded maps. At the same time, Wittstock [242] proved his decomposition theorem, which veriﬁes this conjecture. Independently, Haagerup [106] proved directly that a unital homomorphism of a C ∗ -algebra into B(H) is similar to a ∗-homomorphism if and only if it is completely bounded, and moreover, that there exists a similarity S such that S −1 · S is equal to the cb norm of the homomorphism. It is not clear if this sharp equality can be obtained by combining the results of Hadwin and Wittstock. Haagerup applied his result to obtain the analogous characterization of inner derivations (Corollary 9.10). The result on inner derivations had been obtained earlier by Christensen [54]. In [159], the extension theorem for completely bounded maps into B(H) and the generalization of Stinespring’s representation theorem were proven in order to extend the techniques of Hadwin and Wittstock to operator algebras. It was proven that a unital homomorphism of an operator algebra into B(H) is similar to a completely contractive homomorphism if and only if it is completely bounded. The prime motivation was to prove that an operator is similar to a contraction if and only if it is completely polynomially bounded (Theorem 9.11), which had been conjectured by Arveson. The extension theorem had been obtained earlier by Wittstock [243]. The fact that the similarity could be chosen such that S · S −1 is equal to the cb norm of the homomorphism for a general operator algebra (Theorem 9.1) was obtained later [160]. The key new technique can be found in a paper of Holbrook [120]. The problem of characterizing the operators that are similar to contractions has been considered by Sz.-Nagy. After obtaining his characterization of the operators that are similar to unitaries, he conjectured that an operator T was

132

Chapter 9. Completely Bounded Homomorphisms

similar to a contraction if and only if T n was uniformly bounded for all positive integers n. Such an operator is called power-bounded. Foguel [98] (see also Halmos [110]) gave an example of a power-bounded operator that is not similar to a contraction, and Halmos conjectured that the right condition was polynomially bounded.

Exercises 9.1 Verify the claim of Theorem 9.1, that |·| = |·|n . 9.2 Let A be an algebra (not necessarily unital), and let ρ: A → B(H) be a homomorphism. If ρ(A)H is dense in H, then ρ is called nondegenerate. Show that if A is unital, then ρ is nondegenerate if and only if ρ(1) = 1. 9.3 Let P be in B(H) such that P 2 = P. (i) Show that relative to some decomposition of H,

1 X P= . 0 0 (ii) Show that if S=

1 X , 0 1

then S P S −1 is an orthogonal projection. (iii) Deﬁne ρ: C ⊕ C → B(H) via ρ(λ1 , λ2 ) = λ1 P + λ2 (1 − P), and show that ρ is a completely bounded homomorphism, but that in general, ρcb < S −1 · S. (iv) Compute ρcb . (v) Can you construct an operator R such that RR −1 = ρcb and R P R −1 is an orthogonal projection? 9.4 Let B be a C ∗ -algebra with unit, let A ⊆ B be a subalgebra that does not contain the identity of B, and let ρ: A → B(H) be a homomorphism. Set A1 = {a + λ1: a ∈ A, λ ∈ C}, and deﬁne ρ1 : A1 → B(H) by ρ1 (a + λ1) = ρ(a) + λ · 1H . Prove that ρ1 is a completely bounded homomorphism if and only if ρ is a completely bounded homomorphism. 9.5 Let A be a unital operator algebra, and let ρ: A → B(H) be a bounded homomorphism with ρ(1) = I . (i) Show that if there exists x and y in H such that ρ(A)x = H and ρ(A)∗ y = H, then ρ is completely bounded. (ii) Show that if there exists x1 , . . . , xn and y1 , . . . , ym in H such that ρ(A)x1 + · · · + ρ(A)xn = H and ρ(A)∗ y1 + · · · ρ(A)∗ ym = H, then ρ is completely bounded.

Exercises

133

A subalgebra B of B(H) with a vector x such that Bx = H is called a strictly cyclic algebra. When there exists a ﬁnite set of vectors such that Bx1 + · · · + Bxn = H, then B is called strictly multicyclic. 9.6 Prove that if T is a n × n matrix and X is a K -spectral set for T , then X is a complete (n K )-spectral set for T . Characterize those matrices for which D− is a K -spectral set for some K in terms of their Jordan form. 9.7 (Sz.-Nagy) This exercise gives a more direct proof of Corollary 9.4. Let T be an invertible operator on H such that T n ≤ M for all integers n, and let glim be a Banach generalized limit [34]. (i) Show that x, y1 = glimT n x, T n y deﬁnes a new inner product on H and that (1/M 2 )x, x ≤ x, x1 ≤ M 2 x, x. (ii) Show that T is a unitary transformation on (H, , 1 ). (iii) Prove that there exists a similarity S on H, with S −1 · S ≤ M 2 , such that S −1 T S is a unitary. 9.8 Prove that a ﬁnite direct sum of operators T1 ⊕ · · · ⊕ Tn is similar to a contraction if and only if each operator in the direct sum is similar to a contraction. Moreover, prove that inf{S · S −1 : S −1 (T1 ⊕ · · · ⊕ Tn )S ≤ 1} is achieved by an operator S that is itself a direct sum. 9.9 Let T1 , T2 , T3 be the operators of Parrott’s example, and let ρ be the contractive homomorphism of P(D3 ) deﬁned by ρ(z i ) = Ti . Prove that ρ is completely bounded. 9.10 (Sz.-Nagy–Foias) An operator T in B(H) is said to belong to class Cρ if there exists a Hilbert space K containing H, and a unitary U on K, such that T n = ρ PH U n H for all positive integers n. Prove that such a T is completely polynomially bounded and that there exists an invertible operator S such that S −1 T S is a contraction with S · S −1 ≤ 2ρ − 1 when ρ ≥ 1. 9.11 Let X be a compact set in the complex plane such that R(X ) is dense in C(∂ X ). Prove that X is a K -spectral set for some operator T if and only if T is similar to a normal operator and σ (T ) is contained in ∂ X . Show that the similarity can always be chosen such that S · S −1 ≤ K 2 . 9.12 Let T be an operator with real spectrum. Prove that T is similar to a self-adjoint operator if and only if the Cayley transform of T, C = (T + i)(T − i)−1 , has the property that for some constant M, C n ≤ M for all integers n.

134

Chapter 9. Completely Bounded Homomorphisms

9.13 Let P be a family of (not necessarily self-adjoint) commuting projections on a Hilbert space H satisfying: (a) If P ∈ P, then (1 − P) ∈ P. (b) If P, Q ∈ P, then P Q ∈ P. (c) If P, Q ∈ P, P Q = 0, then P + Q ∈ P. . (d) P = sup{P: P ∈ P} is ﬁnite. (i) For P, Q ∈ P, set P Q = 1 − P − Q + 2P Q, and show that this is an element of P. We call P Q the symmetric difference of P and Q. (ii) Prove that (P, ) is a commutative group. (iii) Prove that ρ(P) = 2P − 1 deﬁnes a representation of this group and that ρ ≤ 2P. (iv) Prove that there exists a similarity S with S · S −1 ≤ 4P2 such that S −1 P S is a self-adjoint projection for all P in P. (v) Use (iv) to give an alternative proof of the fact that every homomorphism of a commutative, unital C ∗ -algebra is similar to a ∗homomorphism. What estimate can you obtain on S · S −1 ? 9.14 Let A be a ﬁnite-dimensional C ∗ -algebra, and let ρ: A → B(H) be a homomorphism, ρ(1) = 1. Show that ρ is completely bounded and that ρcb ≤ ρ2 . 9.15 Let T be in B(H) with σ (T ) contained in the open unit disk. Prove that ∗k k 1/2 T P −1/2 ≤ 1. P= ∞ k=0 T T is a norm-convergent series with P

Chapter 10 Polynomially Bounded and Power-Bounded Operators

Polynomially bounded and power-bounded operators have played an important role in the development of this area, and there are a number of interesting results, counterexamples, and open questions about these operators. In particular, we will present Foguel’s example [98] of a power-bounded operator and Pisier’s example [191] of a polynomially bounded operator that are not similar to contractions. Recall that an operator T is power-bounded provided that there is a constant M such that T n ≤ M for all n ≥ 0. Clearly, if T = S −1 C S with C a contraction, then T is power-bounded with T n ≤ S −1 S. It is fairly easy to see (Exercise 10.1), by using the Jordan form, that a matrix T ∈ Mn is power-bounded if and only if it is similar to a contraction. Sz.-Nagy [229] proved that the same characterization holds when T is a compact operator. This led naturally to the conjecture that an arbitrary operator is similar to a contraction if and only if it is power-bounded. Foguel provided the ﬁrst example of a power-bounded operator that is not similar to a contraction. Recall that an operator is polynomially bounded provided there is a constant K such that p(T ) ≤ K p∞ for every polynomial p, where the ∞norm is the supremum norm over the unit disk. By von Neumann’s inequality, if T = S −1 C S with C a contraction, then T is polynomially bounded with K = S −1 S. So it was natural to conjecture that an operator is similar to a contraction if and only if it is polynomially bounded. Lebow [136] proved that Foguel’s operator was not polynomially bounded, so that it could not serve as a counterexample to this stronger conjecture. This conjecture, in fact, became one of Halmos’s [111] “ten problems” and remained unsolved for about 25 years until Pisier provided a counterexample. Foguel’s counterexample and Pisier’s counterexample share some common features, and a number of researchers have studied operators in this general

135

136 Chapter 10. Polynomially Bounded and Power-Bounded Operators family, so we shall present some general results about operators of what are often called “Foguel type.” We begin with Sz.-Nagy’s result. The reader should also recall Corollary 9.4. Theorem 10.1 (Sz.-Nagy). Let T be a compact operator on a Hilbert space. Then T is similar to a contraction if and only if T is power-bounded. Proof. Clearly every operator similar to a contraction is power bounded. If T is power bounded, say T n ≤ M, then by the spectral radius formula, we have that r (T ) = limn T n 1/2 ≤ limn M 1/n = 1. Since T is compact, there can be at most a ﬁnite number of points in the spectrum of T on the unit circle. If one encloses these by a curve and integrates the resolvent of T about this curve, then by standard facts from the Riesz functional calculus [68], one obtains an idempotent E that commutes with T such that σ (E T | E(H) ) = σ (T ) ∩ T, σ ((I − E)T |(I −E)(H) ) ⊆ D. If we choose an invertible S such that S −1 E S = P is an orthogonal projection, then P commutes with S −1 T S and hence reduces it. Thus, S −1 T S = T1 ⊕ T2 , where σ (T1 ) = σ (T ) ∩ T, σ (T2 ) ⊆ D, and T1 acts on a ﬁnite-dimensional space. Since S −1 T S is power-bounded, T1 is power-bounded and hence by Exercise 10.1 there exists S1 such that S1−1 T1 S1 is a contraction. By Rota’s theorem (Corollary 9.14) there exists an invertible S2 such that S2−1 T2 S2 is a contraction. Setting R = S(S1 ⊕ S2 ) we have that R −1 T R is a contraction. We now focus on the types of operators arising in Foguel’s and Pisier’s counterexamples. Let SH : 2 (H) → 2 (H) be the forward shift, so that SH ((h 0 , ∗ ((h 0 , h 1 , . . . )) = (h 1 , h 2 , . . . ). h 1 , . . . )) = (0, h 0 , h 1 , . . . ), with adjoint SH 2 Let X = (Ai, j ) ∈ B( (H)) where Ai j ∈ B(H) for i, j ≥ 0. We shall call an operator of the form ∗ SH X on 2 (H) ⊕ 2 (H) F= 0 SH a Foguel operator over H with symbol X . Of special interest will be when X = (Ai+ j ) has the Hankel form, and we call these Foguel–Hankel operators over H with symbol X . To simplify notation we generally write S for SH . It is easily seen that for n ≥ 2, n−1 S ∗n X n n F = S ∗ j X S n−1− j . where X n = n , 0 S j=0 Setting X 0 = 0, X 1 = X , the formula holds for all n ≥ 0.

Chapter 10. Polynomially Bounded and Power-Bounded Operators 137 Note that when X has the Hankel form, then S ∗ X = X S and consequently X n = n X S n−1 = n(Ai+ j+n−1 ). Since S ∗n = S n = 1 for all n ≥ 0, we see that X is power-bounded if and only if supn X n ≤ M for some M. Let P as usual denote the space of polynomials, and deﬁne a linear map δ: P → B(2 (H)) by setting δ(z n ) = X n and extending linearly. When X has the Hankel form, we have δ( p) = X p (S), where p is the derivative of p. We now have that p(S ∗ ) δ( p) , p(F) = 0 p(S n ) and since p(S ∗ ) = p(S n ) = p∞ , F is polynomially bounded if and only if δ is a bounded linear map. Analogously, applying the canonical shufﬂe, we have that F is completely polynomially bounded if and only if δ is a completely bounded linear map. One ﬁnal property that is useful is that δ obeys a certain derivation property. Indeed, δ( pq) ( p · q)(S ∗ ) = ( p · q)(F) = p(F)q(F) 0 ( p · q)(S) q(S ∗ ) δ(q) p(S ∗ ) δ( p) = 0 p(S) 0 q(S) p(S ∗ )q(S ∗ ) p(S ∗ )δ(q) + δ( p)q(S) = 0 p(S)q(S) and we have δ( pq) = p(S ∗ )δ(q) + δ( p)q(S) for any polynomials p and q. We begin with Pisier’s counterexample. Our presentation follows [71]. We start with a family of Foguel–Hankel operators that are easy to analyze. To this end, set H = 2 , and let {E i, j }i,+∞ j=0 denote the usual matrix units regarded as elements of B(2 ). To simplify notation we set E i = E i0 . Since E i∗ E j = δi j E 0 , we see that for any vector h = (c0 , c1 , . . . ) ∈ 2 , the operator ∞ ci E i is bounded and in fact T (h) = i=0 T (h)2 = T (h)∗ T (h) = |ci |2 E 0 = h2 . c¯i c j E i∗ E j =

138 Chapter 10. Polynomially Bounded and Power-Bounded Operators Thus, T : 2 → B(2 ) deﬁnes an isometric linear map that identiﬁes a vector with the inﬁnite column matrix. There is one subtle distinction to be made. For while 2 is a Hilbert space, T (2 ) is an operator space. In particular, given vectors (vi j ) in 2 , we have a well-deﬁned norm for (T (vi j )). To compute its norm one simply substitutes the corresponding column matrix for each vector. We shall encounter the operator space T (2 ) again in Chapter 14, where it is called column Hilbert space. Theorem 10.2. Fix a sequence {an }+∞ n=0 , and consider the Foguel–Hankel oper2 ator F over with symbol X = (ai+ j E i+ j ). Then the following are equivalent: (i) (ii) (iii) (iv)

F is similar to a contraction, F is polynomially bounded, F is power-bounded, 2 supn n( ∞ k=n−1 |ak | ) is ﬁnite.

Proof. Clearly, (i) implies (ii), and (ii) implies (iii). Recall that F is powerbounded if and only if supn X n is ﬁnite, where X n = n X S n−1 = n(ai+ j+n−1 E i+ j+n−1 ). We have that X n∗ X n

=n

2

∞

2

∞

∗ a¯ i+k+n−1 a j+k+n−1 E i+k+n−1 E j+k+n−1

k=0

=n

a¯ i+k+n−1 a j+k+n−1 δi j E 0 .

k=0

Thus, X n∗ X n is a diagonal operator matrix whose diagonal entries are n2

∞

|ai+k+n−1 |2 = n 2

k=0

∞ 2

∞

|ak |2 .

k=i+n−1

2 Thus, X n∗ X n = n k=n−1 |ak | , from which the equivalence of (iii) and (iv) follows. Finally, to prove that (iv) implies (i), we consider the operator matrix Y = ( jai+ j−1 E i+ j−1 ) where we set E −1 = 0. Assume for the moment that Y is bounded. We have that Y S − S ∗ Y = ( j + 1)ai+( j+1)−1 E i+( j+1)−1 − ja(i+1)+ j−1 E (i+1)+ j−1

= (ai+ j E i+ j ) = X.

Chapter 10. Polynomially Bounded and Power-Bounded Operators 139 Thus, if we set R = ( 0I

Y I

R −1 F R =

), then R −1 = ( 0I

=

) and we have

I +Y 0 I ∗ ∗ X −YS + S Y S 0 = . S 0 S

I −Y 0 I S∗ 0

−Y I

S∗ 0

X S

Hence, we have that F is similar to a contraction. It remains to prove that Y is bounded. Computing ∗ Y ∗ Y = (i a¯ i+ j−1 E i+ j−1 )( jai+ j−1 E i+ j−1 ) ∞ ∗ = i j a¯ i+k−1 ak+ j−1 E i+k−1 E k+ j−1 , k=0

we see that Y ∗ Y is a diagonal operator matrix whose diagonal entries are ∞ ∞ 2 2 ∗ 2 2 k=0 i |ai+k−1 | E 0 . Thus, Y Y = supn n k=0 |ak+n−1 | , and so (iv) implies (i). Theorem 10.2 shows that the above family of operators could not possibly provide an example of a polynomially bounded operator that is not similar to a contraction. However, to obtain Pisier’s counterexample we will swap the above sequence of operators {E i } for a different sequence {Wi }. The following propositions will allow us to prove that statements (ii), (iii), and (iv) of Theorem 10.2 are still equivalent for the Foguel–Hankel operators obtained from this new sequence of operators, but their equivalence to (i) will no longer hold. Proposition 10.3. Let H and K be Hilbert spaces, and let : B(H) → B(K) be a bounded linear map. If (Ai+ j ) = A is a bounded Hankel operator on 2 (H), then A = ((Ai+ j )) is a bounded Hankel operator on 2 (K). In fact, A ≤ A. Proof. By the Nehari–Page theorem (Theorem 5.10) there exists a sequence {An }−1 n=−∞ such that +∞ |n| inθ A = sup r e An . r <1 n=−∞ ∞

140 Chapter 10. Polynomially Bounded and Power-Bounded Operators |n| inθ (An )∞ ≤ A. Applying the Nehari– Hence, supr <1 +∞ n=−∞ r e Page theorem again, we have that −1 ∞ r |n| einθ Bn + r |n| einθ (An ) , A = inf sup B r <1 n=−∞ n=0 ∞

where the inﬁmum is over all sequences B A ≤ A.

= {Bn }−1 n=−∞

in B(K). Hence,

The above result shows that even though may not be a completely bounded map, for Hankel matrices it acts completely bounded. Proposition 10.4. Let H and K be Hilbert spaces, and let : B(H) → B(K) be a bounded linear map. Let F be a Foguel–Hankel operator over H with symbol X = (Ai+ j ), and let F be the corresponding Foguel–Hankel operator over K with symbol X = ((Ai+ j )). Then: (i) F power-bounded implies F power-bounded, (ii) F polynomially bounded implies F polynomially bounded. Proof. We only prove (ii); the proof of (i) is similar. Recall that F is polynomially bounded if and only if the map δ: P → B(2 (H))δ( p) = X p (S) is bounded. For the operator F , we need to consider the map δ ( p) = X p (S). But X p (S) = (Bi+ j ) is a Hankel operator matrix, and X p (S) = ((Bi+ j )). Hence, δ ( p) ≤ (Bi+ j ) ≤ δ p∞ . Thus, δ is a bounded map with δ ≤ δ, and hence F is polynomially bounded. The analogous conclusion for completely polynomially bounded operators is false, and that will be the source of our counterexamples. Our construction uses a particular choice of a sequence of operators related to the CAR operators. The reader already familiar with sequences of operators satisfying the CAR and their properties might wish to skip directly to Theorem 10.5 and substitute any CAR sequence for the operators {Wi } occurring in Theorem 10.5. A sequence {Cn }∞ n=0 of operators on a Hilbert space H is said to satisfy the canonical anticommutation relations (CAR) provided (CAR a)

Ci C j + C j Ci = 0

Chapter 10. Polynomially Bounded and Power-Bounded Operators 141 and (CAR b)

Ci C ∗j + C ∗j Ci = δi j I

for all i and j, where δi j denotes the Kronecker delta. ∞ αi Ci . Given such a sequence and h = (α0 , α1 , . . . ) in 2 , set (h) = i=0 This deﬁnes a bounded operator with (h)2 ≤ h2 , since (h) (h)∗ + (h)∗ (h) =

∞

αi α¯ j (C j C ∗j + C ∗j Ci ) =

i, j=0

∞

|αi |2 · I,

i=0

by (CAR b). However, if we let P = (h)∗ (h), Q = (h) (h)∗ , then P = Q and P Q = (h)∗ (h)2 (h)∗ = 0, since (h)2 = 0 by (CAR a). Hence ∞ |αi |2 , and it follows that (h) = h. P = Q = P + Q = i=0 2 Thus, : → B(H) is an isometry, and (2 ) is another operator space, isometrically isomorphic to Hilbert space. To construct a sequence of operators satisfying the CAR, we ﬁrst need a ﬁnite sequence of operators on a ﬁnite-dimensional space that satisfy them. To construct these operators we begin with three 2 × 2 matrices,

1 0 1 0 0 0 . V = , C= , and I2 = 0 −1 0 1 1 0 We have that V 2 = I, C 2 = 0, C ∗ C = E 11 , CC ∗ = E 22 , V C = −C, and C V = C. For 0 ≤ i ≤ n − 1, we deﬁne matrices in M2 ⊗ · · · ⊗ M2 ∼ = M2n by setting Ci = V ⊗(i) ⊗ C ⊗ I2⊗(n−i−1) , where X ⊗(k) denotes the tensor product of X with itself k times. Note that (1) Ci2 = (V 2 )⊗(i) ⊗ (C 2 ) ⊗ I2⊗(n−i−1) = 0, (2) Ci∗ Ci = (V ∗ V )⊗(i) ⊗ (C ∗ C) ⊗ I2⊗(n−i−1) = I2⊗(i) ⊗ E 11 ⊗ Ii⊗(n−i−1) , and (3) Ci Ci∗ = I2⊗(i) ⊗ E 22 ⊗ I2⊗(n−i−1) ,

while for 0 ≤ i < j ≤ n − 1, we have ⊗(n− j−1)

= −C j Ci , (4) Ci C j = I2⊗(i) ⊗ C ⊗ V ⊗( j−i−1) ⊗ C ⊗ I2 ⊗(i) ⊗(n− j−1) ∗ ⊗( j−i−1) ∗ (5) Ci C j = I2 ⊗ C ⊗ V ⊗ C ⊗ I2 = −C ∗j Ci . Now (1) and (4) are seen to imply (CAR a), while (2), (3), and (5) yield (CAR b). n−1 αi Ci . From the above relations Given h = (α0 , . . . , αn−1 ), set (h) = i=0 we have that is an isometry from n-dimensional Hilbert space into M2n , and {Ci } span a Hilbert space.

142 Chapter 10. Polynomially Bounded and Power-Bounded Operators Next we show that {Ci ⊗ Ci } nearly span an 1 -space, that is, given scalars a0 , . . . , an−1 we will show that n−1 n−1 n−1 (6) 12 i=0 |ai | ≤ i=0 ai Ci ⊗ Ci ≤ i=0 |ai |. Now up to a unitary equivalence, we have that Ci ⊗ Ci ∼ = (V ⊗ V )⊗(i) ⊗ (C ⊗ ⊗(n−i−1) in M4 ⊗ · · · ⊗ M4 = M4n . Let e1 , e2 denote the basis for C2 , C) ⊗ I4 and choose wi , 1 ≤ i ≤ n, constants √ of modulus one such that ai w¯ i = |ai |. Then xi = (e1 ⊗ e1 + wi e2 ⊗ e2 )/ 2 will be a unit vector in C4 , and x = n x1 ⊗ · · · ⊗ xn will be a unit vector in C4 . Since V ⊗ V (ei ⊗ ei ) = ei ⊗ ei , C ⊗ C(e1 ⊗ e1 ) = e2 ⊗ e2 , and C ⊗ C(e2 ⊗ e2 ) = 0, we have that

n−1 n−1 n−1 1 1 ai Ci ⊗ Ci x, x = ai w¯ i = |ai |. 2 i=0 2 i=0 i=0 n−1 n−1 ai Ci ⊗ Ci ≥ 12 i=0 |ai |, and since Ci ⊗ Ci = 1, the other Thus, i=0 inequality follows. So far we have only constructed, for each n, a ﬁnite sequence C0 , . . . , Cn−1 of ﬁnite matrices satisfying the CAR. To obtain an inﬁnite sequence one generally uses an inﬁnite tensor product of Hilbert spaces. To avoid discussions of inﬁnite tensor products we use a slightly different approach. Relabel the 2n × 2n matrices constructed above as C0,n , . . . , Cn−1,n , 2n and for i ≥ n set Ci,n = 0. Deﬁne bounded operators on ∞ n=1 ⊕ C by setting ∞ Wi = n=1 ⊕ Ci,n . It is readily veriﬁed that ∞ ∞ (i) i=0 αi Wi 2 = i=0 |αi |2 for any (α0 , α1 , . . . ) in 2 , n−1 n−1 n−1 (ii) 12 i=0 |ai | ≤ i=0 ai Ci,n ⊗ Wi ≤ i=0 |ai | for any (a0 , . . . , an−1 ), (iii) Wi W j + W j Wi = 0, (iv) Wi W j∗ + W j∗ Wi = δi j (I − Pi ),

n where Pi is the ﬁnite-rank projection onto in=1 ⊕ C2 . Thus, although {Wi } do not quite satisfy the CAR, their images in the C ∗ -algebra obtained by quotienting out the ideal of compact operators do satisfy them. Theorem 10.5 (Pisier). Fix a sequence {an }∞ n=0 , and consider the Foguel– ∞ are Hankel operator F over H with symbol X = (ai+ j Wi+ j ), where {Wi }i=0 the operators constructed above. Then the following are equivalent: (i) F is polynomially bounded, (ii) F is power-bounded, 2 (iii) supn n ( ∞ k=n−1 |ak | ) is ﬁnite. 2 2 However, if ∞ k=0 (k + 1) |ak | is inﬁnite, then F is not similar to a contraction.

Chapter 10. Polynomially Bounded and Power-Bounded Operators 143 Proof. Clearly (i) implies (ii). To see that (ii) implies (iii), recall that F is power-bounded if and only if supn X n is ﬁnite, where X n = n X S n−1 = n(ai+ j+n−1 Wi+ j+n−1 ). We have that X n∗ X n + X n X n∗ ∞ ∗ ∗ a¯ i+k+n−1 a j+k+n−1 [Wi+k+n−1 = n2 W j+k+n−1 + Wi+k+n−1 W j+k+n−1 ]. k=0

By (iii), the off-diagonal terms are 0, and hence the norm of this operator matrix is just the largest norm of a diagonal entry. Since each diagonal entry is a sum of positive operators, the supremum occurs when i = 0. Thus, ∞ 2 ∗ ∗ + =n |ak+n−1 | Wk+n−1 Wk+n−1 + Wk+n−1 Wk+n−1 k=0 ∞ ∗ ∗ = sup n 2 |ak+n−1 |2 (Ck+n−1,m Ck+n−1,m + Ck+n−1,m Ck+n−1,m ) k=0 m

X n∗ X n

= n2

X n X n∗

∞

2

|ak+n−1 |2 ,

k=0 ∗ ∗ using the fact that Ck,m Ck,m + Ck,m Ck,m = I for k < m. Hence, supn X n ∗ ∗ 2 ﬁnite implies that supn X n X n + X n X n = supn n 2 ∞ k=0 |ak+n−1 | is ﬁnite. To prove that (iii) implies (i), we use Proposition 10.4. Deﬁne : B(2 ) → ∞ ai,0 Wi . Since the norm of a matrix is larger than the B(H) by ((ai, j )) = i=0 ∞ |ai,0 |2 ≤ (ai j )2 and norm of any column, we have that ((ai, j ))2 = i=0 hence ≤ 1. 2 Assuming supn n ∞ k=0 |ak+n−1 | is ﬁnite, by Theorem 10.2 we have that the Foguel–Hankel operator with symbol (ai+ j E i+ j ) is polynomially bounded, and hence by Proposition 10.4, the Foguel–Hankel operator with symbol (ai+ j (E i+ j )) = (ai+ j Wi+ j ) is polynomially bounded. 2 2 Finally, to see that F is not similar to a contraction when ∞ k=0 (k + 1) |ak | is 2 inﬁnite, it is sufﬁcient to show that the map δ: P → B( (H)) is not completely bounded for this operator. Note that the (0,0) entry of δ(z i+1 ) is (i + 1)ai Wi . Thus, if we can show that the map δ0 : P → B(H), δ0 (1) = 0, δ0 (z i+1 ) = (i + 1) ai Wi , is not completely bounded, then δ will fail to be completely bounded, too. n (i + 1) If we consider the 2n × 2n matrix-valued polynomial P(z) = i=0 n n a¯ i Ci,n z i+1 , then δ0(2 ) (P) = i=0 (i + 1)2 |ai |2 Ci,n ⊗ Wi . By the deﬁnition of

144 Chapter 10. Polynomially Bounded and Power-Bounded Operators Wi and (6) for the operators Ci,n we have that n n 1 (2n ) 2 2 δ (P) ≥ (i + 1) |ai | Ci,n ⊗ Ci,n ≥ (i + 1)2 |ai |2 . 0 i=0 2 i=0 n (i + 1)2 |ai |2 )1/2 . Hence, Since is an isometry, we have that P∞ = ( i=0 n 1 2 2 1/2 δ0 cb ≥ 2 ( i=0 (i + 1) |ai | ) , and letting n tend to inﬁnity we have that δ0 is not completely bounded. The above result leads us to Pisier’s counterexample to Halmos’s conjecture. Corollary 10.6. (Pisier). Let a2k −1 = 2−k and ai = 0 otherwise. Then the Foguel–Hankel operator with symbol (ai+ j Wi+ j ), where the W ’s are deﬁned as above, is polynomially bounded, but not similar to a contraction. ∞ −k 2 2 j Proof. We have that supn (n + 1) ∞ k=n−1 |ak | = sup j 2 k= j (2 ) = 4/3, while ∞ ∞ (k + 1)2 |ak |2 = (2 j )2 (2− j )2 = +∞, k=0

j=0

and so we are done by Theorem 10.5.

For another example, one can set ak = (k + 1)−3/2 . Then 1/2 ∞ √ |ak |2 ≤ 2 sup (n + 1) n

k=n

but ∞ (k + 1)2 |ak |2 = +∞. k=1

The same results as above hold when the sequence {Wn } is replaced by an actual CAR sequence {Cn }. Necessary and sufﬁcient conditions on the sequence {an } for the Foguel– Hankel operator with symbol (ai+ j Ci+ j ) to be similar to a contraction were 2 2 only recently obtained. In [197], it is shown that ∞ k=0 (k + 1) |ak | ﬁnite is a necessary and sufﬁcient condition for similarity to a contraction. Earlier [11] it 2 2+ had been shown that if ∞ |ak |2 is ﬁnite for any k=1 (k + 1) [log(log(k + 1))] > 0, then one has similarity to a contraction. We now turn our attention to the problem of exhibiting a power-bounded operator that is not polynomially bounded. Examples of such operators generally

Chapter 10. Polynomially Bounded and Power-Bounded Operators 145 rely on some nontrivial results about A(D). The work of Peller [174] shows that there exist Foguel–Hankel operators with scalar symbol that are powerbounded but not polynomially bounded. However, partly for historical reasons and partly to minimize the amount of function theory that is needed, we shall present Foguel’s operator [98], which was the ﬁrst known example of an operator that is power-bounded, but not similar to a contraction. Later, Lebow [136] proved that Foguel’s operator is not even polynomially bounded, and we reproduce this result in Theorem 10.9. To understand Foguel’s example, it is best if we start by looking at the Foguel operator S ∗ Er,r , Fr = 0 S where Er,r denotes the standard matrix unit. We have that S ∗n X n,r n Fr = , 0 Sn where X n,r =

n−1

S ∗k Er,r S n−1−k =

k=0

n−1

Er −k,r −n+k+1 ,

k=0

and we have adopted the convention that E i, j denotes the standard matrix units provided that i, j ≥ 0 and is 0 otherwise. ∗ is a matrix all of whose entries are 0 or 1 and that Note that X n,r = X n,r the 1’s occur along part of the antidiagonal where i + j = 2r − n + 1. For 2r +1−n 2 = =r n > 2r + 1, we have X n,r = 0 and, in general, X n,r −n E , . Thus, we see that X n,r ≤ 1 for all n and hence Fr is power-bounded. Given r and t, compute X n,r X n,t =

n

Er −k,r −n+k+1 E t−n++1,t− .

k,=0

We see that each term in this product is 0 unless r + k = t + ≥ n − 1, r ≥ k, and t ≥ . These inequalities imply that 2r ≥ t and 2t ≥ r . Thus, if 2r + 1 < t or 2t + 1 < r then X n,r X n,t = 0. Theorem 10.7 (Foguel). Let {k } be a sequence of integers satisfying 2k + S∗ X 1< k+1 , and let X = ∞ =1 E k ,k . Then the operator F = ( 0 S ) is powerbounded.

146 Chapter 10. Polynomially Bounded and Power-Bounded Operators Proof. We have that F = n

S ∗n 0

Xn , Sn

where

Xn =

∞

X n,k .

=1

Since either 2k + 1 < km or 2km + 1 < k for = m, we have that X n∗ X n =

∞ =1

∗ X n,k X n,k =

+1−n ∞ 2k

E j, j .

=1 j=k −n

Because the intervals {[k − n, 2k + 1 − n]}∞ =1 are all disjoint, we have that X n∗ X n ≤ 1. To prove that the operators of the above type are not polynomially bounded, we shall need a result from function theory. Theorem 10.8. Let {k } be a sequence of integers satisfying 2k + 1 < k+1 , k and given f (z) = ∞ k=0 ak z in A(D), deﬁne ( f ) = (ak1 , ak2 , . . . ). Then 2 : A(D) → is contractive and onto. 2 1/2 = f 2 ≤ f ∞ , is contractive. Consider the Proof. Since ( ∞ k=0 |ak | ) 1 −a¯ matrix B = ( a 1 ). Note that 1 + |a|2 0 ∗ B B= 0 1 + |a|2 and so we have that B = 1 + |a|2 . Fix h = (h 1 , . . . , h n , 0, . . . ), a ﬁnitely supported vector in 2 with h2 = 1, and let ¯ e−ik θ 1 − h for 1 ≤ ≤ n. B (eiθ ) = h eik θ 1 Then by the above calculation, we will have that B ∞ = sup{B (eiθ ): 0 ≤ θ ≤ 2π} = 1 + |h |2 . For each m, 1 ≤ m ≤ n, let Fm (eiθ ) = B1 (eiθ ) · · · Bm (eiθ ); n 2 2 then Fm 2∞ ≤ m =1 (1 + |h | ) ≤ exp( =1 |h | ) = e. By induction (Exercise 10.3) one shows that iθ −iθ (e ) p (e ) g m m Fm (eiθ ) = f m (eiθ ) qm (e−iθ ) where gm , f m , pm , qm are polynomials satisfying

Notes

147

gm (0) = 1, deg gm ≤ km , qm (0) = 1, deg qm ≤ km , deg pm ≤ km , ( f m ) = (h 1 , . . . , h m , 0, . . . ). √ Since f n ≤ Fn ∞ ≤ e, by the linearity of we have that for every ﬁnitely supported h ∈ 2 there exists f ∈ A(D) with ( f ) = h and f ≤ √ eh. Given an arbitrary h ∈ 2 , one can choose a sequence of vectors {h n } with ∞ each h n , ﬁnite support such that ∞ n=1 h n ≤ 2h and√ n=1 h n = h. For choose f n ∈ A(D) with ( f n ) = h n and f n ≤ eh n . Then f = ∞ n=1 f n converges in A(D), and ( f ) = h. (1) (2) (3) (4)

Theorem 10.9 (Lebow). Let {k } be a sequence of integers with 2k + 1 < S∗ X k+1 , and let X = ∞ =1 E k ,k . Then the operator F = ( 0 S ) is powerbounded, but not polynomially bounded. Proof. Let h = (1, 1/2, 1/3, . . . ), which is in 2 . Since 2(2k + 1) + 1 < 2k+1 + 1, we see that the sequence {2k + 1} satisﬁes the hypotheses of k Theorem 10.8. Hence there exists f (z) = ∞ k=0 ak z in A(D) with a2k +1 = 1/. If F were polynomially bounded, then Y = ∞ m=1 am X m would be a bounded X . operator, where X m = ∞ =1 m,k Using the formula for X n,r we see that 0 if n = 2r + 1, X n,r e0 , e0 = 1 if n = 2r + 1. Hence, X m e0 , e0 =

0 if m = 2k + 1 for some ,

1 if m = 2k + 1, for some , and thus Y e0 , e0 = =1 a2k +1 = ∞ =1 1/, which diverges. Thus, Y is not a bounded operator, and consequently, F is not polynomially bounded. ∞

Notes Theorem 10.5 with a CAR sequence of operators in place of the sequence {Wn } follows from Pisier [191]. The proof given here follows that of Davidson and the author [71] closely.

148 Chapter 10. Polynomially Bounded and Power-Bounded Operators Some earlier attempts at ﬁnding a counterexample focused on how large the ratio of the complete polynomial bound to the polynomial bound could be for an n × n matrix, together with obtaining asymptotic estimates of the growth of this ratio as n tended to inﬁnity. See [184] for details of this work. There is an extensive literature on Foguel–Hankel operators over C with many deep connections to function theory. See Peller [174] for the earliest published work on this family of operators, where the question arises of determining precisely for which sets of symbols these operators are, respectively, polynomially bounded or similar to contractions. Bourgain [33] proved that such operators are polynomially bounded if and only if the derivative of the symbol is in BMO. Later, Aleksandrov and Peller [4] proved that such operators are polynomially bounded if and only if they are similar to contractions. This led naturally to the study of Foguel–Hankel operators over general Hilbert spaces. In a remarkable paper Pisier [191] generalized many of the function-theoretic results to obtain operator-valued versions, which allowed him to construct his counterexample to the Halmos’s conjecture. In an earlier unpublished work, Foias and Williams [101] studied general Foguel operators and proved that such an operator is similar to a contraction if and only if it is similar to S ⊕ S ∗ . This result motivated the simpliﬁcation of Pisier’s original proof found by Davidson and the author [71]. The result of Foias and Williams has a homological interpretation as the vanishing of a certain Ext group for modules over the disk algebra. The study of this group of extensions was begun by Carlson and Clark [38, 39], and only later was the connection with this unpublished manuscript realized, resulting in [40]. Further contributions were made by Ferguson ([90], [91], and [93]). There are still many open questions about the computation of these Ext groups and their relations to higher-order Hankel forms. See for example [65]. In particular, the study of these groups over polydisk algebras is related to questions about the relationships between the families of “big” and “small” Hankel operators and generalizations of Nehari’s theorem. See for example [96], [94], and [92].

Exercises 10.1 Let



λ 1 0 ···  ... . . . . . . . . .  .. ..  . . Jλ = 0  .. . . .. . . . 0 ... 0 ... be an elementary Jordan block.

 0 ..  .  0  ∈ Mk  1 λ

Exercises

10.2

10.3

10.4 10.5

149

(i) Show that the (1, k) entry of Jλm is the kth derivative of z m evaluated at λ. Deduce that Jλ is power-bounded if and only if |λ| < 1. (ii) Prove that if |λ| < 1 then Jλ is similar to a contraction. (iii) Let T ∈ Mn be power-bounded. Prove that T is similar to a contraction if and only if T is power-bounded. Let T be an operator on a Hilbert space such that σ (T ) = C1 ∪ C2 with C1 and C2 closed, C1 ⊆ D, and C2 ﬁnite. Prove that T is similar to a contraction if and only if T is power-bounded. (i) Show that g1 , f 1 , p1 , q1 , satisfy (1)–(4) in the proof of Theorem 10.8. (ii) Show that if qm , f m , pm , qm satisfy (1)–(4), then gm+1 , f m+1 , pm+1 , qm+1 satisfy (1)–(4). ˙ A(D)/ Let : A(D) → 2 be the map of Theorem 10.8. Prove that : √ 2 −1 ˙ ≤ 1, ˙ ≤ e. ker → is a Banach space isomorphism with Let C1 , . . . , Cn denote the 2n × 2n CAR matrices, let E 11 , . . . , E n1 be the standard matrix units in Mn , and let (λ1 E 11 + · · · + λn E n1 ) = √ λ1 C1 + · · · + λn Cn , so that is an isometry. Prove that cb ≥ n/2.

Chapter 11 Applications to K-Spectral Sets

In this chapter we apply the results of Chapter 9 to the study of multiply connected K-spectral sets. We show that for a “nice” region X with ﬁnitely many holes it is possible to write down a fairly simple characterization of the family of operators that, up to similarity, have normal ∂ X -dilations. This constitutes a model theory for these operators. In contrast, if X has two or more holes, then it is still an open problem to determine whether or not every operator for which X is a spectral set has a normal ∂ X -dilation, i.e., is a complete spectral set. A further difﬁculty with the theory of spectral sets is that it is quite difﬁcult to determine if a given set is a spectral set for an operator. We will illustrate this difﬁculty in the case that X is an annulus and T is a 2 × 2 matrix. Thus, even if it is eventually determined that the properties of being a spectral set and being a complete spectral set are equivalent, the use of the theory might be limited by the impossibility of recognizing operators to which it could be applied. It is easier to determine when a “nice” set with no holes is a spectral set for an operator. If a compact subset X of the complex plane is simply connected with boundary a Jordan curve, then one can deﬁne an analytic homeomorphism f from X to the closed unit disk. In this case it is easily seen that X is a spectral set for an operator T if and only if f (T ) is a contraction. For this reason, criteria for operators to have a simply connected set as a spectral set are fairly readily available. For many ﬁnitely connected sets, there is a simple criterion for the set to be a K-spectral set for an operator. To develop this criterion, we ﬁrst need to extend the deﬁnition of spectral sets to include closed, possibly unbounded proper subsets of the complex plane. To motivate this extended deﬁnition, suppose that S is an invertible operator and S ≤ R, so that the closed disk of radius R is a spectral set for S. The

150

Chapter 11. Applications to K-Spectral Sets

151

fact that this disk is a spectral set for S only tells us that S ≤ R and loses the information that S is invertible. The statement that will capture both pieces of information about S, when appropriately deﬁned, is that the complement of the open disk of radius R −1 is a spectral set for T = S −1 . Let X be a closed, proper subset of C, and let Xˆ denote the closure of X , regarded as a subset of the Riemann sphere. That is, Xˆ = X when X is compact, and otherwise Xˆ is X together with the point at inﬁnity. We let R(X ) denote the quotients of polynomials with poles off Xˆ , that is, the bounded, rational functions on X with a limit at inﬁnity. We regard R(X ) as a subalgebra of C(∂ Xˆ ), which deﬁnes norms on Mn (R(X )). If X is a closed, proper subset of C, and T ∈ B(H) with σ (T ) ⊆ X , then we still seek a functional calculus, i.e., a homomorphism ρ: R(X ) → B(H), given by ρ( p/q) = p(T )q(T )−1 . We say that X is a (completely) K-spectral set for T , provided that ρ is well deﬁned and ρ ≤ K (respectively, ρcb ≤ K ). We use the term (complete) spectral when K = 1. Let ψ be a linear fractional transformation regarded as a map from the sphere to the sphere, let X be a closed, proper subset of C, and suppose that the pole of ψ lies off X , so that ψ( Xˆ ) = ψ(X )− = Y is a compact set in C. If f is in R(Y ), then f ◦ ψ deﬁnes an element of R(X ), and the map ψ ∗ : R(Y ) → R(X ), given by ψ ∗ ( f ) = f ◦ ψ, deﬁnes a completely isometric algebra isomorphism between these algebras. The following results are immediate. Proposition 11.1. Let X be a closed, proper subset of C, and let ψ be a linear fractional transformation with pole off X . Then X is a (complete) K-spectral set for some operator T if and only if ψ(X )− is a (complete) K-spectral set for ψ(T ). Proposition 11.2. Let T be an operator. Then T is invertible with T −1 ≤ R if and only if {z: |z| ≥ R −1 } is a complete spectral set for T . It is now quite easy to illustrate one of the subtleties involved in the study of ﬁnitely connected spectral sets that is eliminated by the study of K-spectral sets. Let X = {z: R −1 ≤ |z| ≤ R}, R > 1, be a spectral set for some operator T , so that necessarily T ≤ R and T −1 ≤ R. This last statement is equivalent to the statement that X 1 = {z: |z| ≤ R} and X 2 = {z: |z| ≥ R −1 } are both spectral sets for T . Since X = X 1 ∩ X 2 , it is natural to ask: If X 1 and X 2 are spectral sets for T , then is X a spectral set for T ? The answer, as we shall show in a moment, is no. Thus, T ≤ R and T −1 ≤ R is not enough to guarantee that X is a spectral set for T .

152

Chapter 11. Applications to K-Spectral Sets

On the other hand, if X 1 and X 2 are, respectively, (complete) K 1 -spectral and (complete) K 2 -spectral sets for T , then we shall show that X is always a (complete) K-spectral set for T , for some K (Theorem 11.5). To construct an example where X is not spectral for T , let 1 −t 1 t T = . , T −1 = 0 1 0 1 It is not difﬁcult to calculate that, if t = R − R −1 , then T = T −1 = R. Thus, X 1 and X 2 are both spectral sets for T . However, |z − z −1 | ≤√R + R −1 on X , while T − T −1 = 2t = 2(R − √ R −1 ) > R + R −1 for R > 3. Thus, X is not a spectral set for T when R > 3. Computing the largest value of t such that X is a spectral set for T involves a considerable knowledge of the function theory of X . The proof of Theorem 11.5 requires the introduction of the concept of decomposability. Let X be a closed, proper subset of C. We call a collection {X i } of closed, proper subsets of C a decomposition of X provided that X = X i and every f ∈ R(X ) can be written as a uniformly convergent series f = i f i , where each f i is in R(X i ) and i f i ≤ K f , where K is independent of f . The least value of K satisfying the above inequality we call the decomposition constant, relative to the decomposition {X i }. If there is a constant K such that the above inequality holds for all F in Mn (R(X )) and all n, then we say that {X i } is a complete decomposition of X and call the least such K the complete decomposition constant. Of course every set has a trivial decomposition, namely itself. However, we shall see that many sets have more interesting decompositions. Before proceeding, we point out the relevance of the above deﬁnitions. Proposition 11.3. Let X be a closed, proper set in C with (complete) decomposition {X i } and (complete) decomposition constant K . If T is an operator such that each X i is a (complete) K i -spectral set for T and L = supi K i is ﬁnite, then X is a (complete) K L-spectral set for T . The algebra R(X ) is called a Dirichlet algebra if R(X ) + R(X ) is dense in C(∂ Xˆ ). We shall call a set X for which R(X ) is a Dirichlet algebra a D-set. This concept should not be confused with the concept of a Dirichlet set. A set X is called a Dirichlet set if every continuous function on ∂ X has a harmonic extension to the interior of X . For example, an annulus is a Dirichlet set that is not a D-set. Let X be a compact set in C whose boundary consists of n + 1 disjoint, rectiﬁable, simple, closed curves (i.e., Jordan curves). Such a set will be called

Chapter 11. Applications to K-Spectral Sets

153

n denote the open a nice n-holed set. If X is a nice n-holed set, let {Ui }i=0 components of C\X , with U0 the unbounded component, and let X i = C\Ui , so that X i , i = 0, . . . , n, is a closed set with X 0 compact. Note that X = X 0 ∩ n the canonical decomposition of X . X 1 ∩ · · · ∩ X n . We call {X i }i=0

Proposition 11.4. Let X be a nice n-holed set. Then the canonical decomposition of X is a complete decomposition of X . Proof. Let {X j }nj=0 be the canonical decomposition of X , and let j = ∂ X j , with 0 oriented counterclockwise, and j , j = 1, . . . , n, oriented clockwise. / j , set If F ∈ Mk (R(X )), then for z ∈ 1 F(w)(w − z)−1 dw. F j (x) = 2πi j Since F(z) = F0 (z) + · · · + Fn (z) for z in the interior of X , it is not difﬁcult to see that each F j (z) extends to deﬁne a function in Mk (R(X j )), which we still denote by F j (z), and that with this extended deﬁnition, F(z) = F0 (z) + · · · + Fn (z) for all z ∈ X . For i = j, let di, j denote the minimum distance between i and j , and let i denote the length of i times (2π)−1 . For i = j, and z ∈ j , we have that Fi (z) ≤ i di,−1j F. Thus, for z ∈ i , F j (z) ≤ F + j di,−1j F, Fi (z) = F(z) − j=i j=i and so we have that Fi ≤ ci F, where ci is a constant independent of k. Hence, i Fi ≤ (c0 + · · · + cn )F, and we have that the canonical decomposition is a complete decomposition. Theorem 11.5. Let X be a nice n-holed set with canonical decomposition n , and let T ∈ B(H). The following are equivalent: {X i }i=0 (i) X is a (complete) K-spectral set for T , for some K , (ii) each X i is a (complete) K i -spectral set for T . Proof. A straightforward application of Propositions 11.3 and 11.4.

In Chapter 10 we proved that the hypothesis that X is a K-spectral set for T does not imply that X is a complete K -spectral set for T for some K even when X is the unit disk. There is another, more geometric version of Theorem 11.5.

154

Chapter 11. Applications to K-Spectral Sets

Corollary 11.6. Let X be a nice n-holed set, with canonical decomposition n , and let T ∈ B(H). Assume that there exist analytic homeomorphisms {X i }i=0 ˆ f i : X i → D− , i = 0, . . . , n. The following are equivalent: (i) there exists an invertible operator S such that S −1 T S has a normal ∂ X dilation, n such that Si−1 f i (T )Si ≤ 1 for (ii) there exist invertible operators {Si }i=0 i = 0, . . . , n. Proof. By Corollary 9.12, statements (i) and (ii) above are equivalent to state ments (i) and (ii) of Theorem 11.5, respectively. The two results above reduce determining whether or not a nice n-holed n , is a complete K-spectral set for an set, with canonical decomposition {X i }i=0 operator T to a ﬁnite set of conditions. As a consequence of these result we have: Corollary 11.7. Let T be an invertible operator, and assume that there are invertible operators S1 and S2 satisfying S1−1 T S1 ≤ R, S2−1 T −1 S2 ≤ r −1 , with r ≤ R. Then there is an invertible operator S such that S −1 T S ≤ R and S −1 T −1 S ≤ r −1 . Proof. When r < R, this is a direct consequence of Corollary 11.6, by letting X = {z: r ≤ |z| ≤ R}, f 0 (z) = z, and f 1 (z) = z −1 . If r = R, then (T /R)n ≤ S1−1 · S1 , and (T /R)−n ≤ S2−1 · S2 . Hence, by Corollary 9.4, there exist an invertible operator S such that S −1 (T /R)S is a unitary. Thus, S −1 T S ≤ R and S −1 T −1 S ≤ r −1 . Theorem 11.8. Let X be a nice n-holed set with canonical decomposition n and (complete) decomposition constant K . If T ∈ B(H) and each X i {X i }i=0 is a spectral set for T , then X is a (complete) K-spectral set for T . Proof. Since each X i is a D-set, the hypothesis that X i is a spectral set implies that T has a normal ∂ X i -dilation, by Theorem 4.4. That is, X i is a complete spectral set for T . The remainder of the proof is a direct application of Propositions 11.3 and 11.4. Let X = {z: r ≤ |z| ≤ R}, r < R, be an annulus, so that the canonical decomposition of X is X 0 = {z: |z| ≤ R} and X 1 = {z: r ≤ |z|}. We have that X 0 is a spectral set for T if and only if T ≤ R, and X 1 is a spectral set for T

Chapter 11. Applications to K-Spectral Sets

155

if and only if T −1 ≤ r −1 . We have seen in an earlier example that T ≤ R and T −1 ≤ r −1 is not enough, in general, to guarantee that the annulus is a spectral set for T . However, by the above theorem, the annulus will be a complete K-spectral set for T and so, up to conjugation by a similarity, T will have a normal ∂ X -dilation. gives It is not difﬁcult to see from the proof of Proposition 11.4, that 2 + R+r R−r an upper bound on the complete decomposition constant for the canonical decomposition of X . It is interesting to contrast this with [211, Proposition 23], where it is shown that the decomposition constant for this annulus is bounded )1/2 . However, that proof does not appear to generalize to other by 2 + ( R+r R−r sets. Precise values for the decomposition constant and complete decomposition constant are not known, even for an annulus. Neither is it known whether these constants are achieved by operators. That is, for T an operator with T ≤ R, T −1 ≤ r −1 , let ρ: R(X ) → B(H) be the homomorphism given by ρ( f ) = f (T ), where X = {z: r ≤ |z| ≤ R}, and set K = sup{ρ: T ≤ R, T −1 ≤ R −1 } and K c = sup{ρcb : T ≤ R, T −1 ≤ r −1 }. By Proposition 11.3, K and K c will be less than or equal to the decomposition constant and the complete decomposition for X , respectively, but it is not known if these inequalities are strict or are in fact equalities. There are sets besides nice n-holed sets for which these decomposition techniques are valuable. For example, the sets

1 1 X 0 = z: |z| ≤ 1 and z − ≥ 2 2 and

X 1 = z: z +

1 1 ≥ 2 4

are D-sets [68, VI.11.11], and the proof of Proposition 11.4 can be suitably modiﬁed to show that these sets deﬁne a complete decomposition of X = X 0 ∩ X 1 . Consequently, the conclusions of Theorems 11.5 and 11.8 apply to these sets as well. Another approach to ﬁnitely connected regions involves the concept of a hypo-Dirichlet algebra. If X is a compact, Hausdorff space, then any subalgebra A of C(X ) that separates points on X and has the property that the closure of

156

Chapter 11. Applications to K-Spectral Sets

A + A¯ is of ﬁnite codimension n in C(X ) is called a hypo-Dirichlet algebra on X of codimension n. Gamelin [103, Theorem IV.8.3] shows that if X is a compact subset of C whose complement has n bounded components, then R(X ) is a hypo-Dirichlet algebra on ∂ X of codimension m, with m ≤ n. In fact, if points {z j }nj=1 are chosen, one from each bounded component of the complement of X , then the span of the union of R(X ), R(X ), and {ln |z − z j |}nj=1 is dense in C(∂ X ). We shall show that unital, contractive homomorphisms of hypo-Dirichlet algebras are necessarily completely bounded. Lemma 11.9. Let A be a unital C ∗ -algebra, and let S ⊆ A be an operator system of codimension n. Then for every ε > 0, there exists a completely positive map φ: A → S and a positive linear functional s on A such that φ ≤ n + 1 + ε, s ≤ n + ε, and φ(x) = x + s(x) · 1 for x ∈ S. Proof. Let π: A → A/S be the quotient map. It is not difﬁcult to show that there exist self-adjoint linear functionals 1 , 2 , . . . , n on A/S and self-adjoint elements h 1 , h 2 , . . . , h n in A/S that form a basis, such that i = h i = 1 and i (h j ) = δi, j , the Kronecker delta. Let h 1 , h 2 , . . . , h n be self-adjoint elements in A such that π(h i ) = h i and h i ≤ 1 + ε/n. Also, let i = i ◦ π , so that S = {y ∈ A: i (y) = 0, i = 1, 2, . . . , n} with i = 1 and i (h j ) = δi, j . By Exercise 7.14, we can write i = f i − gi , where f i ≤ 1, gi ≤ 1, and f i + gi ≤ 1, with f i and gi positive linear functionals on A. If for x in A, we set γ (x) = x −

n

i (x)h i ,

i=1

then γ is a projection of A onto S. Hence the completely positive map φ: A → A deﬁned by φ(x) = x +

n

gi (x)(h i + h i ) +

i=1

= γ (x) +

n

f i (x)(h i − h i )

i=1 n (gi (x) + f i (x))h i · 1 i=1

has range in S. n If we set s(x) = i=1 ( f i (x) + gi (x))h i , then s ≤

n i=1

f i + gi h i ≤ n + ε,

Chapter 11. Applications to K-Spectral Sets

157

and φ(x) = x + s(x) · 1 for x in S. Finally, since φ is completely positive, we have φ = φ(1) = 1 + s(1) ≤ n + 1 + ε. Theorem 11.10. If A ⊆ C(X ) is a hypo-Dirichlet algebra of codimension n, and ρ: A → B(H) is a unital contraction, then ρcb ≤ 2n + 1. Proof. Let ε > 0, let S be the closure of A + A¯ in C(X ), and let φ: C(X ) → S and s be as in the previous lemma. If we extend ρ to ρ: ˜ S → B(H) by ρ( ˜ f + g¯ ) = ρ( f ) + ρ(g)∗ , then ρ˜ will be positive by Proposition 2.12. Thus, ρ˜ ◦ φ: C(X ) → B(H) is positive and hence completely positive. Finally, for f in A, we have ρ( f ) = ρ˜ ◦ φ( f ) − s( f ) · 1H , so that ρcb ≤ ρ˜ ◦ φcb + scb = ρ˜ ◦ φ(1) + s(1) ≤ 2n + 1 + 2ε, since ρ˜ ◦ φ and s are completely positive.

Corollary 11.11. If A ⊆ C(X ) is a hypo-Dirichlet algebra of codimension n, and ρ: A → B(H) is a unital contractive homomorphism, then ρ is similar to a homomorphism that dilates to C(X ). Furthermore, the similarity S may be chosen such that S · S −1 ≤ 2n + 1. Corollary 11.12. Let X be a compact subset of C such that X is a spectral set for T in B(H). If R(X ) is a hypo-Dirichlet algebra of codimension n on ∂ X , then there exists an invertible operator S in B(H) with S −1 · S ≤ 2n + 1 such that S −1 T S has a normal ∂ X -dilation. When X is a nice n-holed set, it is interesting to contrast Corollary 11.12 with Theorem 11.8. If X is a spectral set for T , then both theorems allow us to deduce that X is a complete K-spectral set for T . Corollary 11.12 gives an upper bound on K of 2n + 1, while the bound on K in Theorem 11.8 comes from the complete decomposition constant for X relative to the canonical decomposition of X . The proof of Proposition 11.4 gives a bound on the complete decomposition constant, but the bound one obtains in this manner is always larger than 2n + 1. However, it should be recalled that the hypotheses of Theorem 11.8 are considerably weaker than those of Corollary 11.12.

158

Chapter 11. Applications to K-Spectral Sets Notes

Most of the results of this chapter were obtained by Douglas and the author in [77]. Exercise 11.3 is from [241].

Exercises 11.1 Let T be in B(H) and let X = {z: Re(z) ≥ 0}. Prove that the following are equivalent: (i) X is a complete spectral set for T . (ii) X is a spectral set for T . / σ (C). (iii) C = (T − 1)(T + 1)−1 is a contraction and 1 ∈ (iv) Re(T ) ≥ 0. (v) (T − 1)∗ (T − 1) ≤ (T + 1)∗ (T + 1). Let X be a closed, proper subset of C, T ∈ B(H), and suppose X is a spectral set for T . We say that T has a normal ∂ X -dilation provided that there exists a Hilbert space K containing H and a bounded normal operator N in B(K), with σ (N ) ⊆ ∂ X , such that f (T ) = PH f (N )|H . 11.2 Let X = {z: Re(z) ≥ 0} be a spectral set for T . (i) Prove that T has a normal ∂ X -dilation if and only if the minimal unitary dilation U of the contraction C = (T − 1)(T + 1)−1 satisﬁes 1∈ / σ (U ). (ii) Give an example of an operator T such that X is a complete spectral set for T , but T has no normal ∂ X -dilation. (iii) Let ρ: R(X ) → B(H) be given by ρ( f ) = f (T ). Show that if X is a complete spectral set for T , then ρ has a C(∂ Xˆ )-dilation. 11.3 Let X be a proper, closed subset of C, λ ∈ Xˆ , and T ∈ B(H) with σ (T ) ⊆ X . Set Rλ (X ) = { f ∈ R(X ): f (λ) = 0}. (i) (Williams) Show that if f (T ) ≤ f for all f ∈ Rλ (X ), then X is a spectral set for T . (ii) Show that if f (T ) ≤ K f for all f ∈ Rλ (X ), then X is a (2K + 1)-spectral set. (iii) Give an example where f (T ) ≤ K f for all f ∈ R, but X is not a K-spectral set. (iv)∗ If f (T ) ≤ f for all f ∈ Mn (Rλ (X )) and all n, then is X a complete spectral set for T ? 11.4 Let X i , i = 1, . . . , n, be disjoint, compact sets, and let X be their union. Show that if X is a K-spectral set for T , then T is similar to T1 ⊕ · · · ⊕ Tn where X i is K i -spectral for Ti .

Chapter 12 Tensor Products and Joint Spectral Sets

In this chapter we outline some of the theory of tensor products of C ∗ -algebras, operator spaces, and operator systems, and apply this theory to multivariable dilation theory. An n-tuple of operators (T1 , . . . , Tn ) is said to doubly commute provided that Ti T j = T j Ti and Ti T j∗ = T j∗ Ti for all i = j. We shall see that, for doubly commuting operators, a natural setting for generalizing the theory of spectral sets from a single-variable theory to a multivariable theory is the theory of tensor products of operator systems. Let A and B be unital C ∗ -algebras. Then their tensor product can be made into a ∗-algebra by setting (a ⊗ b)∗ = a ∗ ⊗ b∗ and extending linearly. By a crossnorm, we mean a norm ·γ on A ⊗ B with the property that a ⊗ bγ = ab for a ∈ A and b ∈ B. By a C ∗ -cross-norm, we mean a cross-norm on A ⊗ B that also satisﬁes the C ∗ -algebra axioms, x yγ ≤ xγ yγ and x ∗ xγ = x2γ = x ∗ γ for x, y ∈ A ⊗ B. The completion of A ⊗ B with respect to a C ∗ -cross-norm γ is a C ∗ -algebra, which we denote by A ⊗γ B. In general, there are many possible C ∗ -cross-norms on A ⊗ B, but there are two that we shall be interested in, the maximal and minimal C ∗ -cross-norms. In order to construct the minimal C ∗ -cross-norm, we recall the theory of tensor products of Hilbert spaces. Suppose that H and K are Hilbert spaces. It is well known that if we set h ⊗ k, h ⊗ k = h, h H k, k K and extend linearly, then we obtain an inner product on H ⊗ K. The completion of H ⊗ K with respect to this inner product is a Hilbert space, which we still denote by H ⊗ K. If T and S are operators on H and K, respectively, then setting (T ⊗sp S)(h ⊗ k) = (T h) ⊗ (Sk) extends to deﬁne a bounded, linear operator on H ⊗ K with T ⊗sp S = T S. It is fairly easy to check that (T1 ⊗sp S1 )(T2 ⊗sp S2 ) = (T1 T2 ) ⊗sp (S1 S2 ) and that (T ⊗sp S)∗ = T ∗ ⊗sp S ∗ . If Ai ⊆ B(Hi ), i = 1, 2, are two C ∗ -subalgebras, then we deﬁne the spatial tensor product, A1 ⊗sp A2 , to be the subspace of B(H ⊗ K) spanned by the 159

160

Chapter 12. Tensor Products and Joint Spectral Sets

operators T ⊗sp S with T ∈ A1 and S ∈ A2 . It can be shown that the map from the tensor product A1 ⊗ A2 to A1 ⊗sp A2 is one-to-one and hence setting Ti ⊗ Si sp = Ti ⊗sp Si deﬁnes a cross-norm. From this it follows easily that ·sp is a C ∗ -cross-norm. See [233] for proofs of these facts. If π1 : A1 → B(H1 ) and π2 : A2 → B(H2 ) are unital ∗-homomorphisms, then we get a unital, ∗-preserving homomorphism π1 ⊗ π2 : A1 ⊗A2 → B(H1 ⊗ H2 ) by setting π1 ⊗ π2 (a ⊗ b) = π1 (a) ⊗sp π2 (b). Thus, if for x ∈ A1 ⊗ A2 we set xmin = sup{π1 ⊗ π2 (x): πi : Ai → B(Hi ) unital, ∗-homomorphism, i = 1, 2}, then we obtain a C ∗ -cross-norm on A1 ⊗ A2 . This norm is called the minimal norm, and the completion of A1 ⊗ A2 in this norm is denoted A1 ⊗min A2 . Note that it has the property that if πi : Ai → B(Hi ), i = 1, 2, are any unital ∗-homomorphism, then π1 ⊗ π2 : A1 ⊗ A2 → B(H1 ⊗ H2 ) can be extended, by continuity, to a unital ∗-homomorphism of A1 ⊗min A2 , denoted π1 ⊗min π2 . The following result is explains the name of this C ∗ -cross-norm. For a proof, see [233, Theorem IV.4.19]. Theorem 12.1 (Takesaki). Let A1 and A2 be unital C ∗ -algebras. If γ is a C ∗ -cross-norm on A1 ⊗ A2 , then xmin ≤ xγ for all x ∈ A1 ⊗ A2 . Corollary 12.2. Let A1 and A2 be unital C ∗ -algebras. If πi : Ai → B(Hi ) are one-to-one, unital ∗-homomorphisms, then for x ∈ A1 ⊗ A2 , xmin = π1 ⊗ π2 (x). Proof. By deﬁnition, xmin ≥ π1 ⊗ π2 (x). But setting xγ = π1 ⊗ π2 (x) deﬁnes a C ∗ -cross-norm, from which the other inequality follows. Thus, when Ai ⊆ B(Hi ) are concrete C ∗ -subalgebras the minimal and spatial C ∗ -cross-norms are equal. Suppose that Bi , i = 1, 2, are unital C ∗ -algebras and that Ai ⊆ Bi , i = 1, 2, are unital C ∗ -subalgebras. For x ∈ A1 ⊗ A2 ⊆ B1 ⊗ B2 , we have two possible deﬁnitions of xmin , depending on whether we view it as an element of A1 ⊗ A2 or of B1 ⊗ B2 . However, if we ﬁx πi : Bi → B(Hi ), i = 1, 2, unital, oneto-one ∗-homomorphisms, then since their restrictions are also unital, one-toone ∗-homomorphisms of Ai , i = 1, 2, by Corollary 12.2 we have xmin = π1 ⊗ π2 (x), independent of whether we regard it as an element of A1 ⊗ A2 or B1 ⊗ B2 . This observation can perhaps best be summarized by saying that the natural inclusion of A1 ⊗ A2 into B1 ⊗ B2 extends to a ∗-isomorphism of A1 ⊗min A2 onto the norm closure of A1 ⊗ A2 in B1 ⊗min B2 . Because of this last fact, the minimal C ∗ -cross-norm is also often called the injective C ∗ -cross-norm.

Chapter 12. Tensor Products and Joint Spectral Sets

161

We use this observation to deﬁne a min norm on tensor products of operator spaces. Suppose that Ai , i = 1, 2, are unital C ∗ -algebras and that Si ⊆ Ai , i = 1, 2, are subspaces. We then deﬁne the min norm on S1 ⊗ S2 to be the restriction of the min norm on A1 ⊗ A2 and let S1 ⊗min S2 denote this operator space. Note that if S1 and S2 are operator systems, then S1 ⊗min S2 is also an operator system. It is important to note that since we are deﬁning S1 ⊗min S2 as a concrete subspace of the C ∗ -algebra A1 ⊗min A2 , it is endowed with a matrix norm and, in the case of operator systems, a matrix order. Theorem 12.3. Let Ai and Bi be unital C ∗ -algebras, let Si ⊆ Ai be subspaces, and let L i : Si → Bi be completely bounded, i = 1, 2. Then the linear map L 1 ⊗ L 2 : S1 ⊗ S2 → B1 ⊗ B2 , given by (L 1 ⊗ L 2 )(a1 ⊗ a2 ) = L 1 (a1 ) ⊗ L 2 (a2 ), deﬁnes a completely bounded map L 1 ⊗min L 2 : S1 ⊗min S2 → B1 ⊗min B2 , with L 1 ⊗min L 2 cb = L 1 cb L 2 cb . If S1 and S2 are operator systems and L 1 and L 2 are completely positive, then L 1 ⊗min L 2 is completely positive. Proof. Let Bi ⊆ B(Hi ), i = 1, 2. If we can show that L 1 ⊗ L 2 : S1 ⊗ S2 → B(H1 ⊗ H2 ) is completely bounded in the min norm with L 1 ⊗ L 2 cb = L 1 cb L 2 cb , then we shall be done by the injectivity of the min norm. By the extension theorem for completely bounded maps, we may extend L i to L˜ i : Ai → B(Hi ) with L˜ i cb = L i cb , i = 1, 2. Now applying the generalized Stinespring representation, we obtain unital ∗-homomorphisms πi : Ai → B(Ki ), and bounded operators Vi : Hi → Ki , Wi : Hi → Ki with Vi Wi = L i cb , such that L˜ i (ai ) = Vi∗ πi (a)Wi , ai ∈ Ai ,

i = 1, 2.

Consider V1 ⊗ V2 : H1 ⊗ H2 → K1 ⊗ K2 , W1 ⊗ W2 : H1 ⊗ H2 → K1 ⊗ K2 , and π1 ⊗min π2 : A1 ⊗min A2 → B(K1 ⊗ K2 ). We have that (V1 ⊗ V2 )∗ (π1 ⊗ π2 )(a1 ⊗ a2 )(W1 ⊗ W2 ) = L˜ 1 (a1 ) ⊗ L˜ 2 (a2 ). Thus, L 1 ⊗ L 2 cb ≤ V1 ⊗ V2 W1 ⊗ W2 = L 1 cb L 2 cb on S1 ⊗ S2 . We leave it to the reader to verify that L 1 cb L 2 cb ≤ L 1 ⊗ L 2 cb . If S1 and S2 are operator systems and L 1 and L 2 are completely positive, then we argue as above, using the extension theorem for completely positive maps. In this case, we ﬁnd that Vi = Wi , i = 1, 2, from which the result follows. Corollary 12.4. Let S1 and S2 be operator spaces, x ∈ S1 ⊗ S2 . Then xmin = sup{L 1 ⊗ L 2 (x): L i : Si → B(Hi ), L i cb ≤ 1, i = 1, 2}.

162

Chapter 12. Tensor Products and Joint Spectral Sets

Furthermore, if S1 and S2 are operator systems, then we may take L i to be completely positive, i = 1, 2. The above corollary shows that the min norm on tensor products of operator spaces is invariant under completely isometric isomorphisms of those operator spaces. Let X be a compact, Hausdorff space and let A be a unital C ∗ -algebra. We let C(X ; A) denote the continuous functions from X into A, equipped with the norm F = sup{F(x): x ∈ X } for F ∈ C(X ; A). It is easy to check that if we deﬁne multiplication, addition, and the ∗-operation pointwise, then C(X ; A) is a C ∗ -algebra. Now deﬁne a ∗-homomorphism from C(X ) ⊗ A into C(X ; A) by n i=1

f i ⊗ ai → F(x) =

n

f i (x)ai .

i=1

n f i ⊗ ai γ = F deﬁnes a C ∗ -crossIt is straightforward that setting i=1 norm on C(X ) ⊗ A. Furthermore, a standard partition-of-unity argument shows that the image of C(X ) ⊗ A is dense in C(X ; A). Thus, the above mapping extends to a ∗-isomorphism of C(X ) ⊗γ A with C(X ; A). We can now show that γ is actually the minimal C ∗ -cross-norm. To see this, note that for each ﬁxed x ∈ X , the map f → f (x) extends to a contractive, linear map on C(X ). Thus by Exercise 12.1, the map f ⊗ a → f (x)a extends to n f i ⊗ ai min ≥ a contractive linear map from C(X ) ⊗min A to A, and so i=1 n i=1 f i (x)ai . This shows that the min C ∗ -cross-norm is greater than the γ C ∗ -cross-norm, and consequently they must be equal. We have shown the following: Proposition 12.5. Let X be a compact, Hausdorff space and let A be a unital C ∗ -algebra. Then the map f ⊗ a → f (x)a extends to a ∗-isomorphism between C(X ) ⊗min A and C(X ; A). We note that in the particular case where A = C(Y ) for a compact Hausdorff space Y , the usual identiﬁcation of a continuous function on X × Y with a continuous function from X into C(Y ) is a ∗-isomorphism of C(X ; C(Y )) and C(X × Y ). Thus, C(X × Y ) is ∗-isomorphic to C(X ) ⊗min C(Y ). If Si ⊆ C(X i ) is a subspace, i = 1, 2, then the min norm on S1 ⊗ S2 is just the norm one obtains by viewing an element of S1 ⊗ S2 as a function on X 1 × X 2 . Furthermore, if S1 and S2 are operator systems, then an element of Mn (S1 ⊗ S2 ) will be positive if and only if it is a positive matrix-valued function on X 1 × X 2 .

Chapter 12. Tensor Products and Joint Spectral Sets

163

We now turn our attention to the max norm. Let A and B be unital C ∗ algebras, and let π1 : A → B(H), π2 : B → B(H) be unital ∗-homomorphisms such that π1 (a)π2 (b) = π2 (b)π1 (a) for all a ∈ A and b ∈ B. We may then deﬁne n π1 (ai )π2 (bi ), a unital ∗-homomorphism π: A ⊗ B → B(H) via π (x) = i=1 n where x = i=1 ai ⊗ bi . Conversely, if we’re given a unital ∗-homomorphism π: A ⊗ B → B(H) and we deﬁne π1 (a) = π(a ⊗ 1), π2 (b) = π (1 ⊗ b), then we obtain a pair of unital ∗-homomorphisms of A and B, respectively, with commuting ranges such that π(a ⊗ b) = π1 (a)π2 (b). We deﬁne xmax = sup{π(x): π : A ⊗ B → B(H) unital ∗-homomorphisms}. Thus, if π1 : A → B(H), π2 : B → B(H), are unital ∗-homomorphisms with commuting ranges, then we have a unital ∗-homomorphism π1 ⊗max π2 : A ⊗max B → B(H) satisfying π1 ⊗max π2 (a ⊗ b) = π1 (a)π2 (b). Proposition 12.6. Let A and B be unital C ∗ -algebras, let x ∈ A ⊗ B, and let γ be a C ∗ -cross-norm on A ⊗ B. Then xγ ≤ xmax . Proof. By the Gelfand–Naimark–Segal theorem, there is a unital ∗homomorphism π: A ⊗γ B → B(H) with xγ = π (x). Since π (x) ≤ xmax by deﬁnition, the proof is complete. If we have C ∗ -algebras Ai ⊆ Bi , i = 1, 2, then the natural inclusion A1 ⊗ A2 ⊆ B1 ⊗ B2 ⊆ B1 ⊗max B2 induces a C ∗ -cross-norm γ on A1 ⊗ A2 . The C ∗ algebra A1 ⊗γ A2 can be identiﬁed with the closure of A1 ⊗ A2 in B1 ⊗max B2 . Thus, by Proposition 12.6, we have a ∗-homomorphism A1 ⊗max A2 → B1 ⊗max B2 , which, in general, can be norm-decreasing. Moreover, if πi : Ai → Bi , i = 1, 2, are onto ∗-homomorphisms, then π1 ⊗max π2 : A1 ⊗max A2 → B1 ⊗max B2 is onto, but not necessarily one-to-one. For these reasons, the max norm is often referred to as the projective C ∗ -crossnorm. Unfortunately, the max norm on tensor products is not compatible with completely bounded maps. Huruya [123] has given an example of a completely bounded map L: A1 → A2 and a C ∗ -algebra B such that the map L ⊗ id: A1 ⊗ B → A2 ⊗ B, deﬁned by L ⊗ id(a ⊗ b) = L(a) ⊗ b, does not even extend to a bounded map from A1 ⊗max B into A2 ⊗max B. However, it is the case that the tensor product of completely positive maps yields a completely positive map on the tensor product in the max norm. This fact is a consequence of the following commutant lifting theorem of Arveson [6].

164

Chapter 12. Tensor Products and Joint Spectral Sets

For a set S ⊆ B(H), let S = {T ∈ B(H): T S = ST for all S ∈ S} denote the commutant of S. Note that S is always an algebra and that when S is selfadjoint, S is a C ∗ -algebra. Theorem 12.7 (Arveson). Let H and K be Hilbert spaces, let B ⊆ B(K) be a C ∗ -algebra containing 1K , and let V : H → K be a bounded linear transformation with BV H norm-dense in K. Then for every T ∈ (V ∗ BV ) , there exists a unique T1 ∈ B such that V T = T1 V . Furthermore, the map T → T1 is a ∗-homomorphism of (V ∗ BV ) onto B ∩ {V V ∗ } . Proof. Let A1 , . . . , An be in B, and let h 1 , . . . , h n be in H. Note that if a T1 with the desired properties exists, then n n Ai V h i = Ai V T h i . (∗) T1 i=1

i=1

Since the vectors appearing in the left-hand side of the above formula are dense in K, this shows that such a T1 , provided that it exists, is necessarily unique. Thus, we need to prove that the above formula yields a well-deﬁned, bounded operator. Note that if P and Q are commuting positive operators and x is a vector, then P Qx, x = P Q 1/2 x, Q 1/2 x ≤ P Qx, x. We have that 2 n n Ai V T h i = T ∗ V ∗ Ai∗ A j V T h j , h i i=1 i, j=1 =

n

T ∗ T V ∗ Ai∗ A j V h j , h i = P Qx, x,

i, j=1

where Q = (V ∗ Ai∗ A j V )i,n j=1 , x = h 1 ⊕ · · · ⊕ h n , and P is the diagonal n × n operator, whose entries are T ∗ T . Thus, P and Q are positive and commute, and so 2 2 n n Ai V T h i ≤ P Qx, x = T 2 · Ai V h i . i=1 i=1 This equation shows that the formula (∗) yields a well-deﬁned, bounded operator with T1 ≤ T . From the formula (∗) it is clear that the map π , given by π (T ) = T1 , is a homomorphism into B . To see that it is a ∗-homomorphism, calculate π (T )A1 V h 1 , A2 V h 2 = V ∗ A∗2 A1 V T h 1 , h 2 = T V ∗ A∗2 A1 V h 1 , h 2 = A1 V h 1 , A2 V T ∗ h 2 = A1 V h 1 , π (T ∗ )A2 V h 2 .

Chapter 12. Tensor Products and Joint Spectral Sets

165

Since linear combinations of vectors of the above form are dense in K, we have that π (T )∗ = π(T ∗ ). To see that T1 is also in the commutant of V V ∗ , observe T1 V V ∗ = V T V ∗ = V (V T ∗ )∗ = V (T1∗ V )∗ = V V ∗ T1 . Finally, to see that π is onto B ∩ {V V ∗ } , let X ∈ B ∩ {V V ∗ } and let V = W ∗ P be the polar decomposition of V ∗ , so that X commutes with P = (V V ∗ )1/2 . Let T = W ∗ X W ; then for A ∈ B, ∗

T V ∗ AV = W ∗ X W W ∗ P A P W = W ∗ X P A P W = W ∗ P A P X W = V ∗ A P W W ∗ X W = V ∗ AV T, so that T is in the commutant of V ∗ BV . Also, V T = P W W ∗ X W = P X W = X P W = X V , so that π(T ) = X . Thus, the ∗-homomorphism π is indeed onto B ∩ {V V ∗ } . Note that T is in the kernel of the above ∗-homomorphism π if and only if V T = 0. Thus, if V has trivial kernel, then π is a ∗-isomorphism. It is also worthwhile to note that even when the map π has a kernel, the map θ (X ) = W ∗ X W deﬁnes a completely positive splitting of π, i.e., π ◦ θ(X ) = X . Theorem 12.8. Let A1 , A2 , and B be unital C ∗ -algebras, and let θi : Ai → B, i = 1, 2, be completely positive maps with commuting ranges. Then there exists a completely positive map θ1 ⊗max θ2 : A1 ⊗max A2 → B with θ1 ⊗max θ2 (a1 ⊗ a2 ) = θ1 (a1 )θ2 (a2 ). Proof. Clearly, we may assume that B = B(H). Let (π1 , V1 , K1 ) be a minimal Stinespring representation of θ1, and let γ1 : (V1∗ π1 (A1 )V1 ) → π1 (A1 ) ∩ {V1 V1∗ } be the ∗-homomorphism of Theorem 12.7. We then have that π1 : A1 → B(K1 ), θ˜ 2 = γ1 ◦ θ2 : A2 → B(K1 ) are completely positive maps with commuting ranges. Let V1∗ = W1∗ P1 be the polar decomposition of V1∗ . By the remarks following Theorem 12.7, θ2 (a2 ) − W1∗ θ˜ 2 (a2 )W1 is in the kernel of γ1 and consequently, V1 (θ2 (a2 ) − W1∗ θ˜ 2 (a2 )W1 ) = 0. Thus, V1 θ2 (a2 ) = V1 W1∗ θ˜ 2 (a2 )W1 = P1 θ˜ 2 (a2 )W1 = θ˜ 2 (a2 )P1 W1 = θ˜ 2 (a2 )V1 , since θ˜ 2 (a2 ) commutes with V1 V1∗ = P12 and hence with P1 . Hence, we have that V1∗ π1 (a1 )θ˜ 2 (a2 )V1 = V1∗ π1 (a1 )V1 θ2 (a2 ) = θ1 (a1 )θ2 (a2 ). Repeating the above argument, we let (π2 , K2 , V2 ) be a minimal Stinespring representation of θ˜ 2 and let γ2 : (V2∗ π2 (A2 )V2 ) → π2 (A2 ) ∩ {V2 V2∗ } . Then

166

Chapter 12. Tensor Products and Joint Spectral Sets

π˜ 1 = γ2 ◦ π1 : A1 → B(K2 ) is a ∗-homomorphism whose range commutes with π2 (A2 ). Also V2∗ π2 (a2 )π˜ 1 (a1 )V2 = θ˜ 2 (a2 )π (a1 ). Finally, by the universal property of the max norm, we have a ∗homomorphism, π: A1 ⊗max A2 → B(K2 ) with π (a1 ⊗ a2 ) = π˜ 1 (a1 )π2 (a2 ). Let V : H → K2 be deﬁned by V = V2 V1 , so that θ : A1 ⊗max A2 → B(H), deﬁned by θ(x) = V ∗ π(x)V , is completely positive. Finally, θ(a1 ⊗ a2 ) = V1∗ V2∗ π˜ 1 (a1 )π2 (a2 )V2 V1 = V1∗ π1 (a1 )θ˜ 2 (a2 )V1 = θ1 (a1 )θ2 (a2 ). Thus, θ1 ⊗max θ2 = θ is the desired completely positive map.

Because of the different properties of the max and min norm, it is important to know when they coincide, that is, when there is a unique C ∗ -cross-norm on A ⊗ B. A C ∗ -algebra A that has the property that the max and min C ∗ cross-norms coincide for every unital C ∗ -algebra B is called nuclear. There is a deep and elegant theory characterizing these C ∗ -algebras. See Lance [135] for an excellent survey. For our current purposes it will be enough to know that commutative C ∗ -algebras are nuclear. It is also valuable to note that Mn is nuclear (Exercise 12.4), so that the norm we deﬁned on Mn (A) = Mn ⊗ A is the unique C ∗ -cross-norm. Proposition 12.9. Let X be a compact Hausdorff space. Then C(X ) is nuclear. Proof. Let B be a unital C ∗ -algebra, and let π1 : C(X ) → B(H), π2 : B → B(H), be ∗-homomorphisms with commuting ranges. It will be sufﬁcient to ﬁx n i=1 f i ⊗ bi in C(X ) ⊗ B and show that its max and min norm coincide. By Proposition 12.5 we have that n n f ⊗ bi = sup f (x)bi : x ∈ X . i=1 i i=1 i min

Let E be the B(H)-valued spectral measure associated with the ∗homomorphism π1 . If B is a Borel set in X , then the projection E(B) will commute with π2 (B), since π2 (B) commutes with π1 (C(X )). This implies that n n π1 ( f i )π2 (bi ) = sup E(B) π1 ( f i )π2 (bi ) , i=1 i=1 n π1 ( f i )π2 (bi ) . (E(X \B) i=1

Chapter 12. Tensor Products and Joint Spectral Sets

167

Consequently, if {Uλ } is any open cover of X , n n π1 ( f i )π2 (bi ) = sup E(Uλ ) π1 ( f i )π2 (bi ) . i=1 λ i=1 Now ﬁx ε > 0, and for each x ∈ X , choose an open neighborhood Ux of x such that | f i (x) − f i (y)| < ε for y ∈ Ux , i = 1, . . . , n. This implies that E(Ux )π1 ( f i ) − f i (x)E(Ux ) < ε. Thus, n (π1 ( f i ) − f i (x))π2 (bi ) < ε(bi + · · · + bn ). E(Ux ) i=1 Since the collection {Uλ } forms an open cover, we have n n π ( f )π (b ) ≤ sup f (x)π2 (bi ): x ∈ X i=1 1 i 2 i i=1 i + ε(b1 + · · · + bn ) n ≤ f i ⊗ bi + ε(b1 + · · · + bn ). i=1 min

Finally, using the facts that ε was arbitrary and that the max norm is n f i ⊗ bi max ≤ the supremum over all such π1 and π2 , we have i=1 n i=1 f i ⊗ bi min , which completes the proof. We are now in a position to discuss some of the applications of the tensor theory to operator theory. Recall that a set of operators {Ti } is said to doubly commute if Ti∗ T j = T j Ti∗ and Ti T j = T j Ti for i = j. This is equivalent to requiring that the C ∗ -algebras generated by each of these operators commutes with the C ∗ -algebra generated by any of the other operators, but does not require that each of these C ∗ -algebras be commutative. n Theorem 12.10 (Sz.-Nagy–Foias). Let {Ti }i=1 be a doubly commuting family of contractions on a Hilbert space H. Then there exists a Hilbert space K containing H as a subspace, and a doubly commuting family of unitary operators n on K, such that {Ui }i=1 T k, k≥0 k1 kn T1 (k1 ) · · · Tn (kn ) = PH U1 · · · Un H , where T (k) = ∗−k T , k < 0. n containMoreover, if K is the smallest reducing subspace for the family {Ui }i=1 n n ing H, then {Ui }i=1 is unique up to unitary equivalence. That is, if {Ui }i=1 and

168

Chapter 12. Tensor Products and Joint Spectral Sets

K are another such set and space, then there is a unitary W : K → K leaving H ﬁxed such that W Ui W ∗ = Ui , i = 1, . . . , n. Proof. First, assume that n = 2. Then we have completely positive maps ¯ = p(Ti ) + q(Ti )∗ , i = 1, 2. Since the θi : C(T) → B(H) deﬁned by θi ( p + q) range of θ1 commutes with the range of θ2 , there is a completely positive map θ1 ⊗max θ2 : C(T) ⊗max C(T) → B(H) satisfying θ1 ⊗max θ2 ( f 1 ⊗ f 2 ) = θ1 ( f 1 )θ2 ( f 2 ). However, we have that C(T) ⊗max C(T) is ∗-isomorphic to C(T × T). Thus, we have a completely positive map θ: C(T2 ) → B(H) with θ(z 1k1 z 2k2 ) = T1 (k1 )T2 (k2 ), where z 1 and z 2 are the coordinate functions on T2 . The result now follows for n = 2 by considering the Stinespring representation of θ. For n > 2, by using the associativity of the tensor product (Exercise 12.7), and arguing as above, one obtains a completely positive map θ: C(Tn ) → B(H) with θ(z 1k1 · · · z nkn ) = T1 (k1 ) · · · Tn (kn ). The above result, for n = 2, is weaker than Ando’s Dilation Theorem, since that result does not assume that the operators doubly commute. For operator systems Si ⊆ Ai , i = 1, 2, we wish to deﬁne a commutative max norm. Because of the projective properties of the max norm, it is not sufﬁcient to just consider the norm induced by the inclusion S1 ⊗ S2 ⊆ A1 ⊗max A2 . Instead we take Theorem 12.8 as our deﬁning property. If θi : Si → B(H), i = 1, 2, are maps with commuting ranges, we always have a well-deﬁned map θ1 ⊗ θ2 : S1 ⊗ S2 → B(H). For (xi, j ) ∈ Mn (S1 ⊗ S2 ), we set (xi, j )max = sup{(θ1 ⊗ θ2 (xi, j )): θ : S → B(H), = 1, 2}, where θ1 and θ2 are unital, completely positive maps with commuting ranges and H is an arbitrary Hilbert space. By considering the direct sum of sufﬁciently many of the maps θ1 ⊗ θ2 , we can obtain a unital map γ : S1 ⊗ S2 → B(H) with the property that (xi, j )max = (γ (xi, j )) for all (xi, j ) ∈ Mn (S1 ⊗ S2 ) and all n. We then deﬁne S1 ⊗max S2 to be the concrete operator system that is the image of γ (S1 ⊗ S2 ). Note that if S1 and S2 were actually C ∗ -algebras, then by Theorem 12.8, the norms on Mn (S1 ⊗ S2 ) as above would correspond to the original deﬁnition. That is, given two C ∗ -algebras, their maximal tensor product as operator systems coincides with their maximal tensor product as C ∗ -algebras. We summarize the properties of S1 ⊗max S2 as follows:

Chapter 12. Tensor Products and Joint Spectral Sets

169

Proposition 12.11. Let S1 and S2 be operator systems, B be a C ∗ -algebra, and θi : Si → B, i = 1, 2, be completely positive maps with commuting ranges. Then: (i) there exists a completely positive map θ1 ⊗max θ2 : S1 ⊗max S2 → B with θ1 ⊗max θ2 (a1 ⊗ a2 ) = θ1 (a1 )θ2 (a2 ); (ii) an element (xi, j ) in Mn (S1 ⊗max S2 ) is positive if and only if the matrix (θ1 ⊗max θ2 (xi, j )) is positive in Mn (B) for every pair θ1 , θ2 of completely positive maps with commuting ranges and every C ∗ -algebra B. Proof. Exercise 12.5.

As with C ∗ -algebras, it is important to know when the min and max norms coincide on operator systems. If A is a nuclear C ∗ -algebra, then for every operator system S, the min and max norms on A ⊗ S coincide (Exercise 12.6). Thus, we deﬁne an operator system S to be nuclear if for every operator system T , the min and max norms coincide on S ⊗ T . We now have all the necessary concepts to discuss joint spectral sets and joint dilations. Let X i ⊆ C be compact, i = 1, . . . , n. We set X = X 1 × · · · × X n and deﬁne ∂d X = ∂ X 1 × · · · × ∂ X n . We let Rd (X ) denote the subalgebra of C(∂d X ) spanned by functions of the form r1 (z 1 ) · · · rn (z n ), ri ∈ R(X i ), and let R(X ) denote the subalgebra of C(∂d X ) consisting of the rational functions, R(X ) = { p(z 1 , . . . , z n )/q(z 1 , . . . , z n ): p, q are polynomials, q = 0 on X }. The algebra Rd (X ) is contained in R(X ), is algebraically isomorphic to the tensor product R(X 1 ) ⊗ · · · ⊗ R(X n ), and, in general, is not dense in R(X ). Suppose that Ti ∈ B(H), X i is a K i -spectral set for Ti , i = 1, . . . , n, the set n commutes, and we have a well-deﬁned homomorphism, ρ: Rd (X ) → {Ti }i=1 B(H), given by ρ(r1 · · · rn ) = r1 (T1 ) · · · rn (Tn ). In this case, we call X a joint K n provided that ρ ≤ K , and a complete joint K -spectral spectral set for {Ti }i=1 set provided that ρcb ≤ K . n on a Hilbert space If there exists a family of commuting normals {Ni }i=1 n K, containing H with σ (Ni ) ⊆ ∂ X i , then we call {Ni }i=1 a joint normal ∂d X n provided that r1 (T1 ) · · · rn (Tn ) = PH r1 (N1 ) · · · rn (Nn )|H for dilation of {Ti }i=1 all ri ∈ R(X i ). The following is immediate. n be a family of commuting operators. Then Proposition 12.12. Let {Ti }i=1 n {Ti }i=1 has a joint normal ∂d X -dilation if and only if X is a complete joint n spectral set for {Ti }i=1 . There exists an invertible S with S −1 · S ≤ K such

170

Chapter 12. Tensor Products and Joint Spectral Sets

n has a joint normal ∂d X -dilation if and only if X is a complete that {S −1 Ti S}i=1 n . joint K -spectral set for {Ti }i=1 n be doubly commuting operators, and let X i be Theorem 12.13. Let {Ti }i=1 a complete spectral set for Ti . If R(X i ) + R(X i ) is dense in C(∂ X i ) for i = n has a joint normal ∂d X -dilation. 1, . . . , n − 1, then {Ti }i=1

Proof. Assume n = 2, and let Si denote the closure of R(X i ) + R(X i ) in C(∂ X i ), so that S1 = C(∂ X 1 ). By hypothesis, we have completely positive maps ¯ = f (Ti ) + g(Ti )∗ , with commuting ranges. θi : Si → B(H) satisfying θi ( f + g) Thus, we have a completely positive map θ1 ⊗max θ2 : S1 ⊗max S2 → B(H). But S1 ⊗max S2 = S1 ⊗min S2 and S1 ⊗min S2 is completely isometrically embedded in C(∂ X 1 ) ⊗min C(∂ X 2 ) = C(∂d X ). Thus, for ( f i, j ) ∈ Mn (Rd (X )), we have (ρ( f i, j )) = (θ1 ⊗max θ2 ( f i, j )) ≤ ( f i, j )max = ( f i, j )min , where ρ: Rd (X ) → B(H) is the homomorphism deﬁned by ρ(r1 (z 1 )r2 (z 2 )) = r1 (T1 )r2 (T2 ). But since ( f i, j )min is just the norm of ( f i, j ) in Mn (C(∂d X )), we have that X is a complete joint spectral set for {T1 , T2 }. The proof for n > 2 is analogous. Lemma 12.14. Let X be a compact Hausdorff space, S ⊆ C(X ) an operator system of codimension n, and T another operator system. Then for (ai, j ) ∈ Mk (S ⊗ T ), one has (ai, j )max ≤ (2n + 1)(ai, j )min . Proof. By Lemma 11.9, for any ε > 0 there is a completely positive map φ: C(X ) → S with the property that for f ∈ S, φ( f ) = f + s( f ) · 1, where s is a positive linear functional with s ≤ n + ε, φ ≤ n + 1 + ε. Let a = k i=1 f i ⊗ bi ∈ S ⊗ T , and note that we have a completely positive map φ ⊗ id: C(X ) ⊗max T → S ⊗max T with φ ⊗ id = φ ⊗ id(1 ⊗ 1) ≤ n + 1 + ε. Note that a can also be regarded as an element of C(X ) ⊗ T . But since C(X ) is nuclear and since the min norm is injective, the max and min norms of a in C(X ) ⊗ T and the min norm of a in S ⊗ T all coincide. Thus, we have that k s( f i )1 ⊗ bi , φ ⊗ id(a)max ≤ (n + 1 + ε)amin . But φ ⊗ id(a) = a + i=1 and so s( f i )bi ≤ (n + 1 + ε)amin + (n + ε)amin . amax ≤ φ ⊗ id(a)max + The last inequality follows by noting that s ⊗min id: S ⊗min T → T satisﬁes s ⊗min id = s. Thus, amax ≤ (2n + 1)amin , since ε was arbitrary.

Chapter 12. Tensor Products and Joint Spectral Sets

171

The proof for matrices is similar and uses the complete positivity of all the maps. n+1 Theorem 12.15. Let {Ti }i=1 be doubly commuting operators with X i a complete spectral set for Ti . If R(X i ) is a hypo-Dirichlet algebra of codimension ki on ∂ X i , i = 1, . . . , n, then there exists a similarity S with S · S −1 ≤ n+1 has a joint, normal ∂d X -dilation. (2k1 + 1) · · · (2kn + 1) such that {S −1 Ti S}i=1

Proof. First assume that n = 1. Let Si = R(X i ) + R(X i ), and let φi : Si → ¯ = f (Ti ) + g(Ti )∗ . Then φ1 and φ2 have comB(H) be deﬁned by φi ( f + g) muting ranges, and so there is a completely positive map φ1 ⊗max φ2 : S1 ⊗max S2 → B(H) and φ1 ⊗max φ2 cb = 1. But by the above lemma, the map θ1 ⊗ θ2 : S1 ⊗ S2 → B(H) will extend to be completely bounded in the min norm, with θ1 ⊗min θ2 ≤ 2k1 + 1. Thus, X is a complete, joint (2k1 + 1)spectral set for {T1 , T2 }, from which the result follows. To argue for an arbitrary n, note that by repeated applications of Proposition 12.11 and the above lemma, the identity map on S1 ⊗ · · · ⊗ Sn extends to a completely bounded map from S1 ⊗min · · · ⊗min Sn to S1 ⊗max · · · ⊗max Sn , with id ⊗ · · · ⊗ idcb ≤ (2k1 + 1) · · · (2kn + 1). The proof is now completed as in the n = 1 case. n Theorem 12.16. Let {Ti }i=1 be doubly commuting operators with X i a spectral set for Ti . If R(X i ) is a hypo-Dirichlet algebra of codimension ki on ∂ X i , then there exists a similarity S with S · S −1 ≤ (2k1 + 1) · · · (2kn + 1) such that n has a joint, normal ∂d X -dilation. {S −1 Ti S}i=1

Proof. Fix ε > 0, and let φi : C(∂ X i ) → Si , where Si is the closure of R(X i ) + R(X i ), and si be as in Lemma 11.9. Since X i is spectral for Ti , we have positive maps θi : Si → B(H) with commuting ranges. Hence αi = θi ◦ φi : C(∂ X i ) → B(H) are completely positive. Set βi = θi ◦ si . Again using the nuclearity of C(∂ X i ), we may form the min tensor of any combination of the maps αi and βi and obtain a completely positive map on C(∂d X ). But for f i in Si , θi ( f i ) = αi ( f i ) − βi ( f i ). Hence on S1 ⊗ · · · ⊗ Sn , θ1 ⊗ · · · ⊗ θn = (α1 − β1 ) ⊗ · · · ⊗ (αn − βn ), which, when the right hand side is expanded, expresses θ1 ⊗ · · · ⊗ θn as a difference of sums of completely positive maps on C(∂d X ). Computing the sum of the norms of each of these maps and letting ε tend to 0 yields the result. We close this chapter with one ﬁnal result. Although the hypotheses of the theorem look quite restrictive, many operators that arise in operator theory

172

Chapter 12. Tensor Products and Joint Spectral Sets

have the property that the C ∗ -algebra that they generate is nuclear. For example, Toeplitz operators, subnormal operators, and essentially normal operators usually generate nuclear C ∗ -algebras. n+1 be doubly commuting operators with X i comTheorem 12.17. Let {Ti }i=1 pletely K i -spectral for Ti . If the C ∗ -algebra generated by each Ti is nuclear, i = 1, . . . , n, then there exists a similarity S with S · S −1 ≤ K 1 · · · K n+1 n+1 such that {S −1 Ti S}i=1 has a joint normal ∂d X -dilation.

Proof. By hypothesis, the map ρi : R(X i ) → B(H) given by ρi ( f ) = f (Ti ) is completely bounded. Extend ρi to θi : C(∂ X i ) → B(H) with θi cb = ρi cb . We may then form θ1 ⊗min · · · ⊗min θn = θ: C(∂d X ) → B(H) ⊗min · · · ⊗min B(H), and we shall have that θ cb ≤ ρ1 cb · · · ρn+1 cb . However, for f ∈ Rd (X ), θ( f ) = ρ1 ⊗ · · · ⊗ ρn+1 ( f ) is in C ∗ (T1 ) ⊗ · · · ⊗ C ∗ (Tn+1 ), and the min norm on this latter algebra is the restriction of the min norm on B(H) ⊗min · · · ⊗min B(H). Hence we have that ρ = ρ1 ⊗ · · · ⊗ ρn+1 : Rd (X ) → C ∗ (T1 ) ⊗min · · · ⊗min C ∗ (Tn+1 ) is completely bounded with ρcb ≤ θcb . Now the min and max norms will agree on this latter tensor product, and the tensor product in the max norm maps completely contractively to the C ∗ -algebra generated by T1 , . . . , Tn+1 . Thus, ˜ f 1 (z 1 ) · · · f n+1 (z n+1 )) = we have that the map ρ: ˜ Rd (X ) → B(H) deﬁned by ρ( f 1 (T1 ) · · · f n+1 (Tn+1 ) is completely bounded, from which the result follows. There is a parallel theory for the maximal tensor product of operator algebras studied in [169]. Given unital operator algebras Ai , i = 1, 2, and unital completely contractive homomorphisms πi : Ai → B(H), i = 1, 2, with commuting ranges, one deﬁnes a homomorphism π : A1 ⊗ A2 → B(H) by setting π(a1 ⊗ a2 ) = π1 (a1 )π2 (a2 ). Conversely, given any unital homomorphism π : A1 ⊗ A2 → B(H), if we set π1 (a1 ) = π(a1 ⊗ 1) and π2 (a2 ) = π (1 ⊗ a2 ), then π1 and π2 are unital homomorphisms of A1 and A2 with commuting ranges. Consequently we call a unital homomorphism π : A1 ⊗ A2 → B(H) admissible provided that π1 and π2 are completely contractive. For (xi j ) ∈ Mn (A1 ⊗ A2 ), we deﬁne (xi j )max = sup{(π(xi j )): π admissible} and let A1 ⊗max A2 denote A1 ⊗ A2 endowed with this family of matrix norms. By considering the direct sum of a sufﬁciently large set of admissible

Exercises

173

homomorphisms, one obtains a completely isometric unital homomorphism π: A1 ⊗max A2 → B(H) for some Hilbert space H. Thus A1 ⊗max A2 can be identiﬁed, completely isometrically isomorphically, with the concrete operator algebra π (A1 ⊗max A2 ). The universal operator algebra for n-tuples of commuting contractions (Pn , ·u ), discussed in Chapter 5, can be identiﬁed with the maximal tensor product P(D) ⊗max · · · ⊗max P(D) (n copies), where P(D) is the operator algebra of polynomials in one variable equipped with the supremum norm over the unit disk. In Chapter 16 we will give an abstract characterization of operator algebras, in analogy with the Gelfand–Naimark–Segal characterization of C ∗ -algebras. This characterization will allow us to discuss abstract operator algebras like A1 ⊗max A2 , even in cases where we may not be able to explicitly exhibit a completely isometric representation as an algebra of operators on a Hilbert space. Notes For a more thorough treatment of tensor products of operator systems, see Choi and Effros [49]. In particular, they obtain an abstract characterization of operator systems, which we shall present in Chapter 13. Lance’s survey article [135] and Takesaki’s text [233] are two excellent sources for a further introduction to tensor products and nuclearity. See Dash [70] and Pott [196] for further results on joint spectral sets. Power and the author [169] prove that for any unital operator algebra A, we have that A ⊗min P(D) = A ⊗max P(D) if and only if A ⊗min Tn = A ⊗max Tn for all n, where Tn denotes the algebra of n × n upper triangular matrices. Exercises 12.1 Let Ai , i = 1, 2, be unital C ∗ -algebras, and let f : A1 → C be a bounded linear functional. Prove that there exists a completely bounded map F: A1 ⊗min A2 → A2 with Fcb = f such that F(a1 ⊗ a2 ) = f (a1 )a2 . If f is positive, prove that F is completely positive. 12.2 Let A and B be unital C ∗ -algebras. Prove that if there exists a constant c such that xmax ≤ cxmin for all x ∈ A ⊗ B, then xmax = xmin . 12.3 Let A and B be unital C ∗ -algebras. Verify the following containments: + + ai, j ⊗ bi, j : (ai, j ) ∈ Mn (A) , (bi, j ) ∈ Mn (B) i, j

⊆ (A ⊗ B) ∩ (A ⊗max B)+ ⊆ (A ⊗ B) ∩ (A ⊗min B)+ .

174

Chapter 12. Tensor Products and Joint Spectral Sets

12.4 Prove that every ﬁnite-dimensional C ∗ -algebra is nuclear. 12.5 Let Si , i = 1, 2, be operator systems, and let B be a unital C ∗ -algebra. (i) Prove that if θi : Si → B, i = 1, 2, are unital, completely positive maps with commuting ranges, then there exists a unital, completely contractive map θ1 ⊗max θ2 : S1 ⊗max S2 → B with θ1 ⊗max θ2 (a ⊗ b) = θ1 (a)θ2 (b). (ii) Deduce that θ1 ⊗max θ2 is also completely positive. (iii) Let θi : Si → B(H), i = 1, 2, be completely positive maps with commuting ranges, and let θi (1) = Pi . Prove that there exist unital, completely positive maps θ˜ i : Si → B(H), i = 1, 2, with commut1/2 1/2 ing ranges such that θi = Pi θ˜ i Pi , i = 1, 2. (iv) Prove Proposition 12.11. 12.6 Let A be a C ∗ -algebra. Prove that if A ⊗max B = A ⊗min B for every C ∗ -algebra B, then A ⊗max S = A ⊗min S for every operator system S. 12.7 Let Si , i = 1, 2, 3, be operator systems, and let γ denote either the min or max norm. (i) Prove that S1 ⊗γ (S2 ⊗γ S3 ) and (S1 ⊗γ S2 ) ⊗γ S3 are completely isometrically isomorphic. (ii) Prove that if S1 and S2 are nuclear, then S1 ⊗max S2 = S1 ⊗min S2 is nuclear. 12.8 (Holbrook–Sz.-Nagy) Let S and T be operators that doubly commute. (i) Prove that if S ∈ Cρ and T ∈ Cσ , then ST ∈ Cρσ . (ii) Prove that w(ST ) ≤ Sw(T ). (iii) Prove that w(ST ) ≤ 2w(S)w(T ), and give an example to show that this inequality is sharp. 12.9 Let S1 and S2 be operator systems, let ( pi j ) ∈ Mn (S1 )+ , and let (qk ) ∈ Mm (S2 )+ . Prove that ( pi j ⊗ qk ) ∈ Mnm (S1 ⊗max S2 )+ . 12.10 Let S1 and S2 be operator systems, and let θi : Si → B(H), i = 1, 2, be unital complete isometries with commuting ranges. Deﬁne θ(S1 ⊗ S2 ) = θ1 (S1 )θ2 (S2 ). Prove that for all n, Mn (S1 ⊗max S2 )+ ⊆ {(xi j ) ∈ Mn (S1 ⊗ S2 ): (θ (xi j )) ≥ 0} ⊆ Mn (S1 ⊗min S2 )+ .

Chapter 13 Abstract Characterizations of Operator Systems and Operator Spaces

The Gelfand–Naimark–Segal theorem gives an abstract characterization of the Banach ∗-algebras that can be represented ∗-isomorphically as C ∗ -subalgebras of B(H) for some Hilbert space H. Thus, the GNS theorem frees us from always having to regard C ∗ -algebras as concrete subalgebras of some Hilbert space. At the same time we may continue to regard them as concrete C ∗ -subalgebras when that might aid us in a proof. For example, proving that the quotient of a concrete C ∗ -subalgebra of some B(H) by a two-sided ideal can again be regarded as a C ∗ -subalgebra of some B(K) would be quite difﬁcult without the GNS theorem. On the other hand deﬁning the norm on Mn (A) and many other constructions are made considerably easier by regarding A as a concrete C ∗ -subalgebra of some B(H). In this chapter we shall develop the Choi–Effros [49] abstract characterization of operator systems and Ruan’s [203] abstract characterization of operator spaces. In analogy with the GNS theory, these characterizations will free us from being forced to regard operator spaces and systems as concrete subspaces of operators. We begin with the theory of abstract operator systems. We wish to characterize operator systems up to complete order isomorphism. To this end, let S be a complex vector space, and assume that there exists a conjugate linear map s → s ∗ on S with (s ∗ )∗ = s for all s in S. We call such a space a ∗vector space. We let Sh = {s ∈ S: s = s ∗ } and note that every element x in S can be written x = h + ik with h = (x + x ∗ )/2 and k = (x − x ∗ )/2i both in Sh . For (xi j ) in Mn (S) we set (xi j )∗ = (x ∗ji ), so that Mn (S) is also a ∗-vector space. Clearly, if we wish S to be an operator system, then for each n we shall need a distinguished cone Cn in Mn (S)h that plays the role of the “positive” operators. Moreover, as n varies, these sets should have certain relationships. For example, if p1 is “positive” in S, then ( p01 00 ) should be “positive” in M2 (S). 175

176

Chapter 13. Abstract Characterizations of Operator Systems

Note that, since S is a vector space, there is a natural action of scalar matrices, of the appropriate sizes, on matrices over S. Namely, for A = (ai j ) an m × n matrix and X = (xi j ) an n × k matrix with entries in S, we deﬁne an m × k matrix over S by A · X = ( n=1 ai xj ). We deﬁne multiplication on the right similarly. Thus, for example, p1 0 ∗ . if A = (1, 0), then Ap1 A = 0 0 These considerations motivate the following deﬁnition. Given a ∗-vector space S, we say that S is matrix ordered provided that: (i) for each n we are given a cone Cn in Mn (S)h , (ii) Cn ∩ (−Cn ) = {0} for all n, (iii) for every n and m and A an n × m matrix, we have that A∗ Cn A ⊆ Cm . We call the collection {Cn } a matrix order on S. We adopt the same terminology for maps between matrix ordered *-vector spaces as for operator spaces. Thus, given two matrix-ordered *-vector spaces S and S with cones Cn and Cn , we call a linear map φ: S → S completely positive provided that (xi j ) ∈ Cn implies that (φ(xi j )) ∈ Cn . Similarly, we call φ a complete order isomorphism provided that φ is invertible with both φ and φ −1 completely positive. Finally, we need to axiomatize the role that 1 plays in an operator system. Let S be a matrix-ordered ∗-vector space. We call e ∈ Sh an order unit for S provided that for every x ∈ Sh there exists a positive real r such that r e + x ∈ C1 . We call an order unit e Archimedean if r e + x ∈ C1 for all r > 0 implies that x ∈ C1 . We call e a matrix order unit provided that   e 0 ··· 0 . 0 . . . . . . ..    In =  .. . . . .  . . . 0 0 ··· 0 e is an order unit for Mn (S) for all n, and an Archimedean matrix order unit provided each In is Archimedean. We begin with a few trivial observations. Since there exists r > 0 such that r e + e ∈ C1 , it follows that e = (r + 1)−1 (r e + e) is in C1 , because C1 is a cone. Similarly, In ∈ Cn . If r e + x ∈ C1 , then for any s ≥ r we have se + x = (s − r )e + (r e + x) ∈ C1 . Finally, if x ∈ Sh , then there exists r > 0 such that r e ± x ∈ C1 and hence x = (r e + x)/2 − (r e − x)/2. Thus, Sh = C1 − C1 , which is the condition needed for C1 to be a full cone in S. Similarly, Cn is a full cone in Mn (S) for all n.

Chapter 13. Abstract Characterizations of Operator Systems

177

Our goal is to prove the following: Theorem 13.1 (Choi–Effros). If S is a matrix-ordered ∗-vector space with an Archimedean matrix order unit e, then there exists a Hilbert space H, an operator system S1 ⊆ B(H), and a complete order isomorphism ϕ: S → S1 with ϕ(e) = IH . Conversely, every concrete operator system S is a matrixordered ∗-vector space with an Archimedean matrix order unit e = I H . Before proving this theorem we need two preliminary results. Recall the correspondence between linear functionals s: Mn (S) → C and linear maps φ: S → Mn of Chapter 6. We omit the factor of n used in Chapter 6. Thus, given φ, we deﬁne sφ : Mn (S) → C via sφ ((xi j )) =

n

φ(xi, j )e j , ei ,

i, j=1

where {e1 , . . . , en } is the canonical basis for Cn . Conversely, given s: Mn (S) → C, we deﬁne φs : S → Mn via φs (x) = (si j (x)), where si j (x) = s(x ⊗ E i, j ) and E i, j are the canonical matrix units for Mn . These two operations are mutual inverses, i.e., φ(sφ ) = φ and s(φs ) = s. Proposition 13.2. Let S be a matrix ordered ∗-vector space, let s: Mn (S) → C, and let φ: S → Mn with φ = φs . Then the following are equivalent: (i) s(Cn ) ≥ 0, (ii) φ: S → Mn is n-positive, (iii) φ: S → Mn is completely positive. Proof. The proof is similar to the proof of Theorem 6.1. It is routine to verify that (iii) implies (ii) and that (ii) implies (i). So we prove that (i) implies (iii). First note that, if h = (β1 , . . . , βn ) and k = (α1 , . . . , αn ), then φ(x)h t , k t = s((α¯ i xβ j )) = s(k ∗ xh). Thus, if X = (xi j ) ∈ Cm and v1t , . . . , vmt ∈ Cn , then m

m

φ(xi j )v tj , vit = s(vi∗ xi j v j ) = s(A∗ X A),

i, j=1

where

i, j=1

 v1   A =  ...  vm 

is m × n. Hence, (φ(xi j )) is positive, and so φ is completely positive.

178

Chapter 13. Abstract Characterizations of Operator Systems

Proposition 13.3. Let S be a matrix-ordered ∗-vector space with Archimedean matrix order unit e and for X ∈ Mn (S) set

r In X ∈ C2n . X n = inf r : X ∗ r In Then ·n is a norm on Mn (S), and Cn is a closed subset of Mn (S) in the topology induced by this norm. Proof. We only do the case n = 1. First we prove that x1 ≥ 0. If re x ∈ C2 , x∗ re set A = (10

0 −1);

then

r e −x −x ∗ r e

= A∗

re x A ∈ C2 . x∗ re

Adding these two matrices yields 2r I2 ∈ C2 . Now using I2 ∈ C2 , C2 a cone, and C2 ∩ (−C2 ) = {0} yields r ≥ 0. Next we show that x1 = 0 implies x = 0. If x1 = 0, then re x ∈ C2 for all r > 0. x∗ re Hence

1 ¯ re x (1, λ) = r (1 + |λ|2 )e + λx + (λx)∗ ∈ C1 x∗ re λ

for every complex number λ and any r > 0. By the Archimedean property, λx + (λx)∗ ∈ C1 . Setting λ = ±1 yields x + x ∗ = 0, and setting λ = ±i yields i x + (i x)∗ = 0. Hence, x = 0. Similar tricks show that λx1 = |λ|x1 , x + y1 ≤ x1 + y1 , and x ∗ 1 = x1 , so we leave them to the reader (Exercise 13.5). Finally, we show that C1 is closed. Let xn ∈ C1 , x ∈ S with x − xn 1 → 0. Since xn = xn∗ , we have x = x ∗ . Given any r > 0, choose n so that x − xn 1 < r . Then re x − xn ∈ C2 , x − xn re and picking A = ( 11 ) yields 2r e + 2x − 2xn ∈ C1 . Hence, r e + x ∈ C1 , since xn ∈ C1 and C1 is a cone. Thus, x ∈ C1 by the Archimedean property.

Chapter 13. Abstract Characterizations of Operator Systems

179

We shall refer to the norm given in Proposition 13.3 as the matrix norm induced by the matrix order. Because of the link between norm and order for operators on a Hilbert space, viz., T ≤ 1 if and only if (TI∗ TI ) ≥ 0, we see that any complete order isomorphism φ onto a concrete operator system with φ(e) = I must be a complete isometry of this induced matrix norm. Proof of Theorem 13.1. Assume S is a matrix-ordered ∗-vector space with an Archimedean matrix order unit e. Let Pn = {φ: S → Mn | φ completely posi∞ φ tive, φ(e) = I }, and deﬁne J = ∞ φ∈Pn ⊕ φ: S → φ∈Pn ⊕ Mn , n=1 n=1 where the latter direct sum is in the ∞ sense. Since the latter direct sum is a C ∗ -algebra, to prove that S is completely order-isomorphic to an operator system, it will be enough to prove that J is a complete order isomorphism between S and J (S). To prove this we must show that for (xi j ) ∈ Mn (S), we have (xi j ) ∈ Cn if and only if (J (xi j )) ≥ 0 in the latter C ∗ -algebra. Clearly, by the choice of φ’s, (xi j ) in Cn implies (J (xi j )) ≥ 0. To complete the proof it will be enough to show that if (xi j ) is not in Cn then there exists k and φ ∈ Pk with (φ(xi j )) ≥ 0. Since Cn is closed in the norm topology of Mn (S), by the Krein–Milman theorem (for cones) there exists a linear functional s: Mn (S) → C with s(Cn ) ≥ 0 but s((xi j )) < 0. Consider φs : S → Mn ; then n

φs (xi j )e j , ei = s((xi j )) < 0.

i, j=1

Hence, (φs (xi j )) is not positive. All that remains is to replace φs with a unital completely positive map. Let φs (e) = P ∈ Mn+ . If P is invertible, we may choose A so that A∗ P A = I . Setting ψ(x) = A∗ φs (x)A, we have that ψ(e) = I , n

(ψ(xi j ))A−1 e j , A−1 ei < 0,

i, j=1

and we are done. If P is not invertible, let x ≤ 1. Then (xe∗ xe ) ∈ C2 and hence P φs (x) ≥ 0. φs (x)∗ P From this it follows that if Ph = 0 then φs (x)h = 0. Let Q be the projection onto ker(P)⊥ , so that Qφs (x)Q = φs (x). Let rank(Q) = k, and choose an n × k matrix A and a k × n matrix B so that

180

Chapter 13. Abstract Characterizations of Operator Systems

A∗ P A = Ik , AB = Q. Setting ψ(x) = A∗ φs (x)A, we have that ψ: S → Mk is completely positive with ψ(e) = Ik and n

i, j=1

ψ(xi j )Be j , Bei =

m

φs (xi j )Qe j , Qei =

i, j=1

n

φs (xi j )e j , ei < 0.

i, j=1

Thus, ψ ∈ Pk with (ψ(xi j )) ≥ 0. We leave the proof of the converse statement to the reader.

For an illustration of the uses of the abstract characterization of operator systems, we return to the question examined in Chapter 6 of determining when an operator system S has the property that every positive map with domain S is automatically completely positive. Recall that if T is any other operator system and θ: S → T is a positive map, then θ is necessarily positive on the subset S + ⊗ Mn+ of Mn (S)+ and hence on its closure. Let Cn denote the closure of S + ⊗ Mn+ , with C1 = S + . Clearly, for every n and m, Cn is a cone, Cn ∩ (−Cn ) = (0), and given any m × n matrix B, we have that B ∗ Cm B ⊂ Cn . Thus, the collection {Cn }n≥1 is a (possibly new) matrix order on S. We let S M denote this new matrix-ordered space. We leave it as an exercise (Exercise 13.4) to show that the Archimedean matrix order unit e is also an Archimedean matrix order unit for S M . Hence, S M is an abstract operator system. Clearly, the identity map from S to S M is a positive map. If the identity map from S to S M is completely positive, then Cn = Mn (S)+ for all n, and consequently every positive map with domain S is completely positive. Conversely, if every positive map with domain S is completely positive, then we must have that the identity map from S to S M is completely positive and hence Cn = Mn (S)+ . Thus, we have obtained a new proof of the equivalence of (i) and (iii) in Corollary 6.7. Note that this new proof avoids the need for a separating-functional argument. This essentially happens because the separating-functional argument is built into the proof of the characterization of operator systems. We now turn our attention to the abstract characterization of operator spaces. It will be convenient to consider rectangular, as well as square, matrices. Given a vector space V , we let Mm,n (V ) denote the vector space of m × n matrices over V . We call V a matrix normed space provided that we are given norms ·m,n on Mm,n (V ) such that whenever A = (ai j ) ∈ M p,m , X = (xi j ) ∈ Mm,n (V ), and

Chapter 13. Abstract Characterizations of Operator Systems

181

B = (bi, j ) ∈ Mn,q , then A · X · B p,q ≤ AX m,n B, where A·X ·B =

m

n

aik xk bj

∈ M p,q (V )

k=1 =1

and A, B are the operator norms on M p,m = B(Cm , C p ) and Mn,q = B(Cq , Cn ), respectively. The above axioms guarantee, among other things, that if a matrix is enlarged by adding rows and columns of 0’s, then its norm will remain unchanged (Exercise 13.1). We abbreviate Mn,n (V ) = Mn (V ) and ·n,n = ·n . Let V be a matrix-normed space, and for X ∈ Mm,n (V ), Y ∈ M p,q (V ) deﬁne X ⊕ Y = ( X0 Y0 ) ∈ Mm+ p,n+q (V ), where the 0’s indicate matrices of 0’s of appropriate sizes. We call the matrix norm on V an L ∞ -matrix norm, and call V an L ∞ -matrix-normed space provided that X ⊕ Y m+ p,n+q = max{X m,n , Y p,q } for all m, n, p, q, X, and Y . Note that for any matrix-normed space, max{X m,n , Y p,q } ≤ X ⊕ Y m+ p,n+q , so it sufﬁces to prove the other inequality. Theorem 13.4 (Ruan’s theorem). Let V be a matrix-normed space. Then there exists a Hilbert space H and a complete isometry ϕ: V → B(H) if and only if V is an L ∞ -matrix-normed space. Proof. It is easy to see that every concrete subspace of some B(H) is an L ∞ matrix-normed space. Thus, if V is completely isometric to such a subspace, it must be an L ∞ -matrix-normed space. So assume that V is an L ∞ -matrix-normed space. We shall create an abstract operator system S, i.e., a matrix-ordered ∗-vector space with an Archimedean matrix order unit e, that contains V as a vector subspace and such that the matrix norm induced by the order structure restricts to the matrix norm on V . By Theorem 13.1, S will be completely order-isomorphic to an operator system. But this complete order isomorphism will necessarily be a complete isometry of this induced matrix norm. Thus, if we can create such an “abstract” operator system S, the representation of S as a “concrete” operator system will give us the desired complete isometry of V . We now construct S. The idea is to “abstractly” recreate the operator system SM of Lemma 8.1 that comes from a concrete operator space M. To this end, let V ∗ denote a complex conjugate copy of the vector space V , ¯ ∗ . As a vector i..e, V ∗ = {v ∗ : v ∈ V } with v1∗ + v2∗ = (v1 + v2 )∗ and λv ∗ = ( λv) space, S will be C ⊕ C ⊕ V ⊕ V ∗ , but it is easiest to understand the matrix

182

Chapter 13. Abstract Characterizations of Operator Systems

order if we write

λ v ∗ ∗ : λ, µ ∈ C, v ∈ V , w ∈ V . w∗ µ

S=

We make S a ∗-vector space by setting ∗ ¯ w λ λ v = ¯ v∗ µ w∗ µ so that

r1 v : r , r ∈ R, v ∈ V . 1 2 v ∗ r2

Sh =

We set

C1 =

r1 v 2 , r ≥ 0, v ≤ r r : r 1 2 1 2 . v ∗ r2

This deﬁnition is motivated by the following fact for SM , when M is a concrete r m operator space: For r1 , r2 > 0, (m1∗ r2 ) ≥ 0 if and only if −1/2 −1/2 r1 r1 0 0 r1 m −1/2 −1/2 m ∗ r2 0 r2 0 r2 −1/2 −1/2 r2 m 1 r ≥ 0, = −1/2 −1/2 ∗ 1 1 r1 r2 m −1/2 −1/2

and this latter occurs if and only if r1 r2 m ≤ 1. To deﬁne Cn , note that after a canonical shufﬂe we may write

A X ∗ ∗ Mn (S) = : A, B ∈ Mn , X ∈ Mn (V ), Y ∈ Mn (V ) Y∗ B and that Mn (S)h consists of the above matrices with A = A∗ , B = B ∗ , and Y ∗ = X ∗ = (x ∗ji ). Thus, we deﬁne Cn as the set

P X X∗ Q

: P, Q ≥ 0 and (P + ε I )−1/2 X (Q + ε I )−1/2 ≤ 1 for all real ε > 0 .

We now show that S is matrix-ordered and that e = (10 01) is an Archimedean matrix order unit.

Chapter 13. Abstract Characterizations of Operator Systems

183

To this end, suppose that S = (Si j ) is in Mn (S), that A = (ai j ) ∈ Mm,n , and that λi j xi j Si j = , yi∗j µi j with S ∼ = (YL∗

X M)

Thus, if S ∼ = ( XP∗

after the canonical shufﬂe. Then

AL AX . A·S = ai Sj ∼ = AY ∗ AM X Q)

is in Cn , we must show that A P A∗ AX A∗ AS A∗ ∼ = AX ∗ A∗ AQ A∗

is in Cm . We have AX ∗ A∗ = (AX A∗ )∗ , and it remains to show that for ε > 0, (A P A∗ + ε I )−1/2 (AX A∗ )(AQ A∗ + ε I )−1/2 ≤ 1. Since (A P A∗ + ε I )−1/2 (AX A∗ )(AQ A∗ + ε I )−1/2 = (A P A∗ + ε I )−1/2 A(P + δ I )+1/2 × (P + δ I )−1/2 X (Q + δ I )−1/2 × (Q + δ I )1/2 · A∗ (AQ A∗ + ε I )−1/2 , it will be enough to show that for δ > 0 suitably chosen, each factor in brackets has norm less than 1. By hypothesis (P + δ I )−1/2 X (Q + δ I )−1/2 ≤ 1 for any δ > 0. Now (A P A∗ + ε I )−1/2 A(P + δ I )1/2 ≤ 1 if and only if A(P + δ I )A∗ ≤ A P A∗ + ε I, which holds for any δ ≤ ε/A∗ A. Similarly, it can be seen that (Q + δ I )1/2 A∗ (AQ A∗ + ε I )−1/2 ≤ 1 for any δ ≤ ε/A∗ A. Thus, for any A ∈ Mm,n we have that ACn A∗ ⊆ Cm . We use this fact to see that Cn is a cone. Clearly, tCn ⊆ Cn for t ≥ 0. If Si , i = 1, 2, are in Cn , then the L ∞ condition implies that S1 ⊕ S2 is in C2n . Taking A = (In , In ) we have that A(S1 ⊕ S2 )A∗ = S1 + S2 ∈ Cn , and so Cn is a cone. We claim that e = (10 01) is a matrix order unit. We must show that if ( XH∗ KX ) is in Mn (S)h , then there exists r ≥ 0 so that In 0 H X + r X∗ K 0 In

184

Chapter 13. Abstract Characterizations of Operator Systems

is in Cn . Let r1 = max{H , K }, r2 = X ; then r 1 In + H 0

0 r 1 In + K

∈ Cn and

r 2 In X∗

X r 2 In

∈ Cn ,

and so r = r1 + r2 sufﬁces. The fact that e is Archimedean is trivial from the use of ε > 0 in the deﬁnition of Cn . Finally, if we embed V into S via v → (00 v0), then this mapping is easily seen to be a complete isometry from V into S equipped with the matrix norm induced by the order. Thus, we have produced the desired “abstract” operator system S, and the proof is complete. In Chapter 14 we shall take a closer look at abstract operator spaces, focusing on a few important examples and classes of operator spaces that serve as an introduction to the ﬁeld.

Notes Since the time that Ruan’s theorem ﬁrst appeared, it has been customary to identify L ∞ -matrix-normed spaces and operator spaces as one and the same thing. In fact, many books start by deﬁning the term “operator space” to mean an L ∞ -matrix-normed space and then prove that such objects can be represented completely isometrically as spaces of operators on some Hilbert space. From this viewpoint, a theory of operator spaces can be developed that parallels much of the classical theory of Banach spaces. Two excellent such texts are Pisier’s [193] An Introduction to the Theory of Operator Spaces (Cambridge University Press) and Effros and Ruan’s [88] Operator Spaces (Oxford University Press). In keeping with standard usage, from this point on we shall generally refer to L ∞ -matrix-normed spaces simply as operator spaces. Occasionally, we shall use the terms abstract operator space to emphasize that we are dealing with an L ∞ -matrix-normed space, and concrete operator space for an actual subspace of the space of operators on a Hilbert space. Similarly, we shall identify matrix-ordered ∗-vector spaces with Archimedean matrix order units and operator systems, using the terms abstract operator system and concrete operator system when we wish to emphasize the difference.

Exercises

185

Exercises 13.1 Let V be a matrix-normed space, let X ∈ Mm,n (V ), and let Y ∈ M p,q (V ) be a matrix obtained from X by introducing ﬁnitely many rows and columns of 0’s. Prove that X m,n = Y p,q . 13.2 Let V be a vector space, and assume that we are given a sequence of norms ·n on Mn (V ) satisfying: (i) AX Bn ≤ AX n B for X ∈ Mn (V ), A ∈ Mn , and B ∈ Mn ; (ii) for X ∈ Mn (V ), X ⊕ 0m+n = X n , where 0 denotes an m × m matrix of 0’s. For X ∈ Mm,n (V ) set X m,n = Xˆ , where = max{m, n} and Xˆ is the matrix obtained by adding sufﬁciently many rows or columns to X to make it square. Prove that (V, ·m,n ) is a matrix-normed space. These alternate axioms are often given as the axioms for a matrix-normed space, and consequently no mention is made of the norms of rectangular matrices. 13.3 Let V be an operator space, let W be a closed subspace, and let q: V → V /W denote the quotient map q(v) = v + W . Prove that if we deﬁne norms on Mn,m (V /W ) by setting (q(vi j ))n,m = inf{(vi, j + wi, j )n,m : wi, j ∈ W }, then V /W is an operator space. 13.4 Prove that S M is an operator system. 13.5 Verify the claims of Proposition 13.3.

Chapter 14 An Operator Space Bestiary

In the last chapter we obtained an abstract characterization of operator spaces that allows us to deﬁne these spaces without a concrete representation. This result has had a tremendous impact and has led to the development of a general theory of operator spaces that parallels in some ways the development of the theory of Banach spaces. In this chapter we give the reader a brief introduction to some of the basics of this theory, focusing on some of the more important operator spaces that we will encounter in later chapters. We have already encountered one example of the power of this axiomatic characterization. In Exercise 13.3, it was shown that if V is an operator space and W ⊆ V a closed subspace, then V /W is an operator space, where the matrix norm structure on V /W comes from the identiﬁcation Mm,n (V /W ) = Mm,n (V )/Mm,n (W ). Yet in most concrete situations it is difﬁcult to actually exhibit a concrete completely isometric representation of V /W as operators on a Hilbert space. The ﬁrst natural question in the area is as follows: If V is, initially, just a normed space, then is it always possible to assign norms ·m,n to Mm,n (V ) for all m and n in such a fashion that V becomes an operator space? The answer to this question is yes. Consider the dual space V ∗ of V , and let V1∗ denote its unit ball, equipped with the weak∗ topology. The continuous functions on this compact, Hausdorff space C(V1∗ ) constitute an abelian C ∗ -algebra, and the map j: V → C(V1∗ ) deﬁned by j(v)( f ) = f (v) for f ∈ V1∗ is a linear isometry. Since subspaces of C ∗ -algebras are operator spaces, identifying V with j(V ) induces a particular family of norms on Mm,n (V ) that makes V an operator space. Thus, for (vi j ) in Mm,n (V ) we have that (vi j )m,n = ( j(vi j ))m,n = sup{( f (vi j )) Mm,n : f ∈ V1∗ },

186

Chapter 14. An Operator Space Bestiary

187

where ( f (vi j )) Mm,n indicates the norm of the scalar m × n matrix ( f (vi j )) viewed as a linear transformation from n-dimensional Hilbert space to m-dimensional Hilbert space. Clearly, we need a way to distinguish between the ordinary normed space V and the operator space j(V ). It is standard to let MIN(V ) denote the operator space j(V ) ⊆ C(V1∗ ). This operator space is often called the minimal operator space of V , for reasons made apparent below. For an arbitrary m × n matrix over V , we simply write (vi j )MIN(V ) to denote its norm in Mm,n (MIN(V )). Theorem 14.1. Let V be a normed space, H be a Hilbert space, and let ϕ: V → B(H) be an isometric map. Then for (vi j ) in Mm,n (V ) we have m n m 2 ≤ 1, (i) (vi j )MIN(V ) = sup{ i=1 j=1 αi vi j β j V : i=1 |αi | n 2 j=1 |β j | ≤ 1}, where the supremum is over all αi , β j in C satisfying the inequalities; (ii) (ϕ(vi j )) B (H(n) ,H(m) ) ≥ (vi j )MIN(V ) . Proof. We have that (vi j )MIN(V ) = sup{( f (vi j )) Mm,n : f ∈ V1∗ } n m m n ∗ 2 2 = sup f αi vi j β j : f ∈ V1 , |αi | ≤ 1, |β j | ≤ 1 i=1 j=1 i=1 j=1 m m n n 2 2 = sup αβ v : |αi | ≤ 1, |β j | ≤ 1 , i=1 j=1 i j i j i=1 j=1 V

which proves (i). To see (ii), note that (ϕ(vi j )) B (H(n) ,H(m) )      β1 k α1 h    ..   ..  ≥ sup (ϕ(vi j ))  .  ,  .  :   αm h βn k h, k ∈ H, h = 1, k = 1,

m i=1

|αi |2 ≤ 1,

n j=1

|β j |2 ≤ 1

    

188

Chapter 14. An Operator Space Bestiary n m = sup ϕ αi β j vi j h, k : i=1 j=1 h, k ∈ H, h = 1, k = 1,  m n  = sup ϕ αi β j vi j  i=1 j=1

: B(H)

m

|αi | ≤ 1, 2

n

i=1

j=1

m

n

|αi |2 ≤ 1,

i=1

|β j | ≤ 1 2

 

|β j |2 ≤ 1

j=1



= (vi j )MIN(V ), by (i) and the fact that ϕ is an isometry.

The beauty of (i) is that it gives a formula “internal” to V to compute MIN(V ). Now that we have a “minimal” way to represent a normed space as an operator space, it is not difﬁcult to create a “maximal” representation. We shall denote this operator space by MAX(V ). To deﬁne MAX(V ), given (vi j ) in Mm,n (V ), we set (vi j )MAX(V ) = sup{(ϕ(vi j )) B(H(n) ,H(m) ) : ϕ: V → B(H) isometric}, where the supremum is taken over all Hilbert spaces H and all linear isometries ϕ: V → B(H). This operator space is called the maximal operator space of V . The construction of MIN(V ) guarantees that the collection of such isometries is nonempty and hence vV = vMAX(V ) . Since every “abstract” operator space, can be represented as a “concrete” operator space, any L ∞ -matrix-norm structure that we could endow V with must be smaller than MAX(V ). The following gives a more concrete realization of these matrix norms. Theorem 14.2. Let V be a normed space and let (vi j ) be in Mm,n (V ). Then (vi j )MAX(V ) = inf{AB: A ∈ Mm,k , B ∈ Mk,n , yi ∈ V , yi ≤ 1, 1 ≤ i ≤ k, and (vi j ) = A Diag(y1 , . . . , yk )B}, where Diag(y1 , . . . , yk ) represents the k × k diagonal matrix with entries y1 , . . . , yk , and the inﬁmum is taken over all ways to represent (vi j ) as such a product. Proof. Let (vi j )m,n denote the inﬁmum appearing on the right hand side of the above equation. If ϕ: V → B(H ) is any linear map and (vi j ) = A Diag(y1 , . . . , yk )B is any such factorization, then (ϕ(vi j )) = A Diag(ϕ(y1 ), . . . , ϕ(yk ))B,

Chapter 14. An Operator Space Bestiary

189

and consequently, (ϕ(vi j )) B (H(n) ,H(m) ) ≤ AB· max{ϕ(y1 ), . . . , ϕ(yk )}. Hence, (vi j )MAX(V ) ≤ (vi j )m,n . To prove the other inequality, since MAX(V ) is the largest of all possible L ∞ -matrix norms, it will sufﬁce to prove that V equipped with {·m,n } is an L ∞ -matrix-normed space. First, to verify that ·m,n is a norm on Mm,n (V ), let (vi j ) and (wi j ) in Mm,n (V ) be given, and ﬁx > 0. Then there exist integers k1 , k2 , elements yi ≤ 1, 1 ≤ i ≤ k1 , x j ≤ 1, 1 ≤ j ≤ k2 , of V , and matrices A1 ∈ Mm,k1 , B1 ∈ Mk1 ,n , A2 ∈ Mm,k2 , B2 ∈ Mk2 ,n such that (vi j ) = A1 Diag(y1 , . . . , yk1 )B1 and (wi j ) = A2 Diag(x1 , . . . , xk2 )B2 with A1 B1 ≤ (vi j )m,n + and A B2 ≤ (wi j )m,n + . Replacing A1 , B1 by r A1 , r −1 B1 with r = √2 B1 /A1 , we may assume that A1 = B1 and similarly that A2 = B2 . We have that B1 (vi j ) + (wi j ) = (A1 , A2 ) Diag(y1 , . . . , yk1 , x1 , . . . , xk2 ) B2 and hence (vi j ) + (wi j )m,n

B 1 ≤ (A1 , A2 ) B2 ≤ A1 A∗1 + A2 A∗2 1/2 B1∗ B1 + B2∗ B2 1/2 ≤ (A1 2 + A2 2 )1/2 (B1 2 + B2 2 )1/2 ≤ (vi j )m,n + (wi j )m,n + 2 .

Since was arbitrary, the triangle inequality follows. It is clear that λ(vi j )m,n = |λ|(vi j )m,n and hence ·m,n is a norm. If (vi j ) = A Diag(y1 , . . . , yk )B, and C and D are scalar matrices of appropriate sizes, then C(vi j )D = (C A) Diag(y1 , . . . , yk )(B D). Thus, C(vi j )D ≤ (CD)(AB), and taking the inﬁmum over all such representations of (vi j ) yields C(vi j )D ≤ CD(vi j )m,n . Hence (V, {·m,n }) is a matrix norm. Finally, to see the L ∞ condition, let (vi j ) ∈ Mm,n (V ), (wi j ) ∈ M p,q (V ), and factor (vi j ) = A1 Diag(y1 , . . . , yk1 )B1 , (wi j ) = A2 Diag(x1 , . . . , xk2 )B2 with A1 = B1 ≤ (vi j )m,n + and A2 = B2 ≤ (wi j ) p,q + . Then

190

Chapter 14. An Operator Space Bestiary

(vi j ) ⊕ (wi j ) = (A1 ⊕ A2 ) Diag(y1 , . . . , yk1 , x1 , . . . , xk2 )(B1 ⊕ B2 ) and hence (vi j ) ⊕ (wi j )m+ p,n+q ≤ max{(vi j )m,n + , (wi j ) p,q + }, from which the L ∞ condition follows. This completes the proof.

Note that while the original deﬁnition of MAX(V ) was “extrinsic” in the sense that it required looking at all representations of V as operators on a Hilbert space, Theorem 14.2 gives an “intrinsic” characterization of this norm. Also, since the original formula for MAX(V ) involved a supremum while Theorem 14.2 involves an inﬁmum, we have methods for obtaining upper and lower bounds on norms in MAX(V ). The main idea of the above proof was to “guess” a factorization formula for MAX(V ) and then to verify it by showing that it satisﬁed Ruan’s axioms. Such factorization formulas will appear frequently in later chapters and are one of the main tools of this ﬁeld. Since every operator space structure on a normed space V lies between MIN(V ) and MAX(V ), it is natural to wonder to what extent MIN(V ) and MAX(V ) differ. One way to measure this difference is to regard the identity map on V as a map i: MIN(V ) → MAX(V ) and attempt to compute its cb norm. This number is denoted α(V ), that is, (vi j )MAX(V ) : (vi j ) ∈ Mm,n (V ), m, n arbitrary . α(V ) = icb = sup (vi j )MIN(V ) Thus, α(V ) = 1 if and only if the identity map is a complete isometry from MIN(V ) to MAX(V ), which is equivalent to there existing a unique operator space structure on V . Currently, there are only two normed spaces known 1 2 with α(V ) = 1, namely, ∞ 2 and 2 , which denote C equipped with the norms (x1 , x2 )∞ = max{|x1 |, |x2 |} and (x1 , x2 )1 = |x1 | + |x2 |, respectively. It is still unknown if there exist any other normed spaces V for which α(V ) = 1, but if they exist, they must be of dimension 2, since it is known that α(V ) > 1 whenever dim(V ) ≥ 3. See [163] and [193] for these results. Exact values of α(V ) are still unknown for most important ﬁnite-dimensional normed spaces. While an extensive development of this constant is beyond our current interests, we do record some of the important facts taken from [163], p [164], and [193]. We adopt the convention that n denotes Cn equipped with the p-norm. Theorem 14.3. Let V be a normed space and V ∗ denote its dual space. Then (i) α(V ) = α(V ∗ ),

(ii) (iii) (iv) (v) (vi)

Chapter 14. An Operator Space Bestiary 191 √ if V is n-dimensional, then n/2 ≤ α(V ) ≤ n, if V is inﬁnite-dimensional, √ then α(V ) = +∞, 2 ) ≤ n/ 2, while (n + 1)/2 ≤ α( 2n ) for n odd and for all n, α( n √ 2 2 n + 2n/2 ≤ α( n ) for n even, √ √ for all n ≥ 2, n/2 ≤ α( 1n ) ≤ n − 1, if dim(V ) = dim(W ), then α(V ) ≤ d(V, W )α(W ), where d(V, W ) = inf{T T −1 : T : V → W is invertible}.

Theorem 14.3(iii) implies that if dim(V ) = +∞, then there always exists a Hilbert space H and a bounded linear map ϕ: MIN(V ) → B(H) that is not completely bounded. When dim(V ) < +∞ and ϕ: MIN(V ) → B(H), then (Exercise 14.2) ϕcb ≤ α(V )ϕ, so that α(V ) gives the best estimate of the ratio of the cb norm to the norm. √ Combining the lower bounds on α( 2n ) with the fact that d( 2n , V ) ≤ n for every normed space V of dimension n and applying Theorem 14.3(vi) yields the fact, cited earlier, that α(V ) > 1 for dim(V ) ≥ 3. One difﬁculty in computing α(V ) is that concrete representations of MAX(V ) are known for few normed spaces V . For example, no concrete representation is known for MAX( 2n ), which is the main obstruction to computing an exact value for α( 2n ). One of the rare exceptions is 1n . If Fn denotes the free group on n generators {u 1 , . . . , u n } and Cu∗ (Fn ) is the universal C ∗ -algebra of this group, then (Exercise 14.3) it can be shown that the map ϕ: MAX( 1n ) → Cu∗ (Fn ) given by ϕ((λ1 , . . . , λn )) = λ1 u 1 + · · · + λn u n is a complete isometry. By comparison, if Tn = Rn /Zn denotes the standard n-torus, so that C(Tn ) is the abelian C ∗ -algebra generated by {z 1 , . . . , z n } – where z k = eitk , 1 ≤ k ≤ n, is unitary – then the map ψ: MIN( 1n ) → C(Tn ) given by ψ((λ1 , . . . , λn )) = λ1 z 1 + · · · + λn z n is a complete isometry (Exercise 14.4). Thus, MIN( 1n ) is represented by the “universal” n-tuple for commuting unitaries, while MAX( 1n ) is represented by the “universal” n-tuple for noncommuting unitaries. We shall now take a closer look at MAX(V ) and MIN(V ) for a general ﬁnite-dimensional normed space V . To this end, let B ⊆ Cn be the closed unit ball of some norm on Cn ; equivalently, let B be a bounded, closed, absorbing, absolutely convex set. If we set x = inf{t: t −1 x ∈ B, t > 0}, then · is the norm on Cn and B is the (closed) unit ball of this norm. We shall

192

Chapter 14. An Operator Space Bestiary

write (Cn , B) when we want to indicate this normed space. Thus, for example −

2 (Cn , Dn ) = ∞ and (Cn , B− n n ) = n ,

where Bn = {(λ1 , . . . , λn ): |λ1 |2 + · · · + |λn |2 < 1} is the (complex) Euclidean ball. The polar of B is the set B ∗ = {y ∈ Cn : |y · x| ≤ 1 for all x ∈ B}, where y · x is the usual dot product. The polar of B is easily seen to be the closed unit ball of the dual of (Cn , B), so that (Cn , B)∗ = (Cn , B ∗ ). Let {e1 , . . . , en } denote the standard basis for Cn . Given (vi j ) ∈ M p,q (Cn ), let A = (vi j · e ), 1 ≤ ≤ n, denote the scalar p × q matrices obtained by taking the coefﬁcients of the vectors (vi j ) with respect to the th basis vector. With respect to the identiﬁcation M p,q (Cn ) = M p,q ⊗ Cn we have that (vi j ) = A1 ⊗ e1 + · · · + An ⊗ en . In this manner we also identify M p,q (Cn ) with n-tuples of p × q scalar matrices, and also write (vi j ) = (A1 , . . . , An ). Now to specify an operator space structure on (Cn , B) it is enough to deﬁne the unit ball of M p,q (Cn ) for all p and q and then verify that the resulting families of norms satisfy Ruan’s axioms. Thus for a clearer understanding of MIN((Cn , B)) and MAX((Cn , B)) we wish to identify the n-tuples of matrices (A1 , . . . , An ) that belong to their unit balls. Theorem 14.4. Let B ⊆ Cn be the unit ball of some norm on Cn , let V = (Cn , B) and let A1 , . . . , An be p × q matrices. We have: (i) A1 ⊗ e1 + · · · + An ⊗ en MIN(V ) ≤ 1 if and only if µ1 A1 + · · · + µn An M p,q ≤ 1 for all (µ1 , . . . , µn ) ∈ B ∗ , (ii) A1 ⊗ e1 + · · · + An ⊗ en MAX(V ) ≤ 1 if and only if A1 ⊗ C1 + · · · + An ⊗ Cn M p,q (B(H)) ≤ 1 for all Hilbert spaces H and n-tuples (C1 , . . . , Cn ) of operators on H satisfying λ1 C1 + · · · + λn Cn ≤ 1 for all (λ1 , . . . , λn ) ∈ B. Proof. We ﬁrst prove (ii). Note that ϕ: V → B(H) is contractive if and only if the n-tuple Ci = ϕ(ei ) satisﬁes λ1 C1 + · · · + λn Cn ≤ 1 for all (λ1 , . . . , λn ) ∈ B. Now if (vi j ) = A1 ⊗ e1 + · · · + An ⊗ en , then (ϕ(vi j )) = A1 ⊗ ϕ(e1 ) + · · · + An ⊗ ϕ(en ) = A1 ⊗ C1 + · · · + An ⊗ Cn . Since A1 ⊗ e1 + · · · + An ⊗ en MAX(V ) ≤ 1 if and only if A1 ⊗ ϕ(e1 ) + · · · + An ⊗ ϕ(en ) ≤ 1 for all Hilbert spaces H and ϕ: V → B(H) contractive, (ii) follows. To prove (i), recall that MIN(V ) is the operator space structure obtained by the embedding j: V → C(V1∗ ). But the unit ball of V ∗ is just the polar B ∗ and

Chapter 14. An Operator Space Bestiary

193

j(ei )((µ1 , . . . , µn )) = µi . Hence, A1 ⊗ e1 + · · · + An en MIN(V ) = A1 ⊗ j(e1 ) + · · · + An ⊗ j(en ) M p,q (C(B ∗ )) = sup{µ1 A1 + · · · + µn An M p,q : (µ1 , . . . , µn ) ∈ B ∗ }, and (i) follows.

As an application of these concepts we consider a generalization of the classical Schwarz inequality from complex analysis. Recall that the Schwarz inequality says that if f : D → D is analytic with f (0) = 0, then | f (0)| ≤ 1. Moreover, by considering functions of the form f (z) = αz we see that the set of possible values of f (0) is exactly D− . Suppose now that G ⊆ Cn is the open unit ball of some norm on Cn . Given F = ( f i j ): G → M p,q analytic, we let D F(0) denote the n-tuple of p × q matrices ∂ f i j (0) ∂ f i j (0) ,..., . D F(0) = ∂z 1 ∂z n We are interested in determining the set of all possible n-tuples D F(0) for F: G → ball(M p,q ) with F(0) = 0. A description of these n-tuples can be found in most books on several complex variables; see [205] for example. But those descriptions should be compared with the following for clarity and the ease with which it can be recalled. Proposition 14.5 (Generalized Schwarz lemma). Let G ⊆ Cn be the open unit ball of some norm; let V = (Cn , G − ). There exists F: G → ball(M p,q ) analytic, F(0) = 0, such that D F(0) = (A1 , . . . , An ) if and only if A1 ⊗ e1 + · · · + An ⊗ en MIN(V ∗ ) ≤ 1. Proof. Assume A1 ⊗ e1 + · · · + An ⊗ en MIN(V ∗ ) ≤ 1. Since V ∗ = (Cn , G ∗ ) and (G ∗ )∗ = G − , by Theorem 14.4 (i) we have A1 λ1 + · · · + An λn ≤ 1 for all (λ1 , . . . , λn ) ∈ G − . Deﬁne F: G → M p,q by F((z 1 , . . . , z n )) = A1 z 1 + · · · + An z n . Then F(G) ⊆ ball(M p,q ), F(0) = 0, and D F(0) = (A1 , . . . , An ). Conversely, assume F: G → ball(M p,q ) is analytic, F(0) = 0, and D F(0) = (A1 , . . . , An ). Fix unit vectors h ∈ Cq , k ∈ C p and a point (λ1 , . . . , λn ) ∈ G, and deﬁne f : D → C via f (z) = F(λ1 z, . . . , λn z)h, k. Since f (0) = 0 and f (D) ⊆ D, by Schwarz’s lemma we have that | f (0)| = |(A1 λ1 + · · · + An λn )h, k| ≤ 1. Since h and k are arbitrary unit vectors, we have that A1 λ1 + · · · + An λn ≤ 1 for all (λ1 , . . . , λn ) ∈ G. Thus, A1 ⊗ e1 + · · · + An ⊗ en MIN(V ∗ ) ≤ 1 by another application of Theorem 14.4 (i).

194

Chapter 14. An Operator Space Bestiary

The following is yet another way, in operator space language, to interpret the generalized Schwarz lemma. Let H ∞ (G) denote the space of bounded analytic functions on G. If for ( f i j ) ∈ M p,q (H ∞ (G)) we set ( f i j )∞ = sup{( f i j (z)) M p,q : z ∈ G}, then H ∞ (G) is an operator space. Let H0∞ (G) = { f ∈ H ∞ (G): f (0) = 0} and ∞ (G) = { f ∈ H ∞ (G): f (0) = 0, D f (0) = 0} denote the corresponding opH00 ∞ (G) erator subspaces. The map D: H0∞ (G) → Cn , f → D f (0) has kernel H00 ∞ ∞ n ˙ and so induces a quotient map D: H0 (G)/H00 (G) → C . It is now easy to see that Proposition 14.5 is equivalent to the statement that D˙ is a complete ∞ (G) and MIN(V ∗ ). isometry between the quotient operator space H0∞ (G)/H00 ∞ (G) is often called the cotangent space of G at 0. The space H0∞ (G)/H00 Thus, the generalized Schwarz lemma is simply one way to describe what the natural operator space structure is on the cotangent space. The generalized Schwarz lemma and Theorem 14.4 also make it possible to generalize the phenomenon in Parrott’s example. Theorem 14.6. Let G ⊆ Cn be the open unit ball of a norm on Cn , and let V = (Cn , G − ). Then there exists a Hilbert space H and a unital contractive homomorphism ρ: H ∞ (G) → B(H) with ρcb ≥ α(V ). Proof. Given a Hilbert space H and operators A1 , . . . , An ∈ B(H), we deﬁne ρ: H ∞ (G) → B(H ⊕ H) via f (0) f (0) A1 + · · · + ∂∂z An f (0)I ∂∂z 1 n ρ( f ) = . 0 f (0)I Using the product rule, it is easily checked that ρ is a homomorphism. We claim that if A1 µ1 + · · · + An µn ≤ 1 for all (µ1 , . . . , µn ) ∈ G ∗ , then ρ ≤ 1. To prove this claim, ﬁrst consider the case where f (0) = 0, f ∞ < 1. Then D f (0) = (µ1 , . . . , µn ) ∈ G ∗ by Proposition 14.5, and so ρ( f ) = A1 µ1 + · · · + An µn ≤ 1. For arbitrary f ∈ H ∞ (G), with f ∞ < 1, let ¯ −1 . It is easily seen that σ (ρ( f )) = α = f (0) and let ϕα (z) = (z − α)(1 − αz) f (0) = α and that ϕα (ρ( f )) = (ρ( f ) − α I )(I − αρ( ¯ f ))−1 = ρ(( f − α1)(1 − α¯ f )) = ρ(ϕα ◦ f ). Since ϕα ◦ f (0) = 0 and ϕα ◦ f ∞ ≤ 1, we have that ϕα (ρ( f )) ≤ 1. Clearly, ρ( f ) = ϕ−α (ϕα (ρ( f ))), and hence we may apply von Neumann’s inequality to deduce that ρ( f ) = ϕ−α (ϕα (ρ( f ))) ≤ 1.

Chapter 14. An Operator Space Bestiary

195

Applying Theorem 14.4 (i), we see that A1 ⊗ e1 + · · · + An ⊗ en MIN(V ) ≤ 1 implies ρ ≤ 1. Given any p × q matrices B1 , . . . , Bn such that B1 λ1 + · · · + Bn λn ≤ 1, we have that F((z 1 , . . . , z n )) = B1 z 1 + · · · + Bn z n ∈ M p,q (H ∞ (G)) and F ≤ 1. Since D F(0) = (B1 , . . . , Bn ), setting F = ( f i j ), we have 0 A1 ⊗ B1 + · · · + An ⊗ Bn (ρ( f i j )) = . 0 0 Thus, ρcb ≤ A1 ⊗ B1 + · · · + An ⊗ Bn . Taking the supremum over all such (B1 , . . . , Bn ) and applying Theorem 14.4 (ii), we obtain ρcb ≥ A1 ⊗ e1 + · · · + An ⊗ en MAX(V ) . Since this holds for any A1 ⊗ e1 + · · · + An ⊗ en MIN(V ) ≤ 1, we may take a sequence of such A’s, with A1(k) ⊗ e1 + · · · + A(k) n ⊗ en MIN(V ) (k) ≤ 1 and limk A(k) ⊗ e + · · · + A ⊗ e = α(V ). Finally, setting 1 n MAX(V ) n 1 ∞ (k) A = k=1 ⊕A and using these operators to deﬁne a homomorphism ρ, we have that ρ = 1 while ρcb ≥ α(V ). Applying this construction to the polydisk G = Dn , we obtain an n-tuple of commuting contractions that induce a contractive homomorphism ρ of H ∞ (Dn ) √ n/2. When n = 3, we ﬁnd that this construction eswith ρcb ≥ α( ∞ n )≥ sentially reduces to the example of Parrott presented in Chapter 7. As mentioned earlier, α(V ) > 1 whenever dim(V ) ≥ 3. Thus, for every G ⊆ Cn , n ≥ 3 as above, there exists a unital contractive homomorphism ρ: H ∞ (G) → B(H) with ρcb > 1. If we let P(G) ⊂ H ∞ (G) denote the subalgebra generated by the polynomials in the coordinate functions, then it is not hard to see (Exercise 14.10) that the analogue of Theorem 14.6 holds with P(G) in place of H ∞ (G). Thus, for such a set G there always exists an n-tuple of commuting operators such that G − is a joint spectral set for the n-tuple, but not a complete spectral set. These considerations show how special Ando’s dilation theorem is, since the bidisk is one of the very few possible domains for which contractive homomorphisms can be completely contractive. An important feature of normed spaces is that the space B(V, W ) of bounded linear maps between two normed spaces V and W is again a normed space and that, in particular, the dual V ∗ = B(V, C) of a normed space V is again a normed space. This feature is shared by operator spaces and completely bounded maps. That is, if E and F are operator spaces, then there is a natural way to make CB(E, F), and hence E ∗ , into operator spaces.

196

Chapter 14. An Operator Space Bestiary

To see how this is done suppose that (ϕi j ) ∈ Mm,n (CB(E, F)). We identify (ϕi j ) with a map : E → Mm,n (F) by setting (e) = (ϕi j (e)), and deﬁne a norm on Mm,n (CB(E, F)) by setting (ϕi j ) Mm,n (CB(E,F)) = CB(E,Mm,n (F)) . In short, we identify Mm,n (CB(E, F)) = CB(E, Mm,n (F)). It is not difﬁcult to see that (ϕi j ) Mm,n (C B(E,F)) = sup{(ϕi j (ek )) Mmp,nq (F) : (ek ) ∈ M p,q (E), (ek ) ≤ 1}, where the matrix (ϕi j (ek )) can be interpreted as the p × q matrix of m × n matrices   (e11 ) . . . (e1q )  ..  ..  . , . (e p1 ) . . . (e pq ) or as the m × n matrix of p × q matrices   (ϕ11 (ek )) . . . (ϕ1n (ek ))   .. .. ,  . . (ϕm1 (ek )) . . . (ϕmn (ek )) since these two matrices differ by a canonical shufﬂe. Proposition 14.7 (Ruan). Let E and F be operator spaces, and equip matrices over CB(E, F) with the norms obtained by the identiﬁcations Mm,n (CB(E, F)) = CB(E, Mm,n (F)). Then CB(E, F) is an operator space. Proof. We leave it to the reader (Exercise 14.5) to verify that CB(E, F) satisﬁes the axioms of an L ∞ -matrix-normed space. One special case of the above result is when F = C. Recall that for every linear ϕ: E → C we have ϕ = ϕcb . Thus, E ∗ = B(E, C) = CB(E, C), isometrically. Hence, by the above result there is a natural way to endow the Banach space dual of an operator space with a matrix-normed structure such that the dual of an operator space is again an operator space. In short, we identify Mm,n (E ∗ ) = CB(E, Mm,n ). Thus, if (ϕi j ) ∈ Mm,n (E ∗ ), then we identify it with the map : E → Mm,n given by (e) = (ϕi j (e)) and set (ϕi j ) = cb .

Chapter 14. An Operator Space Bestiary

197

In the future we shall refer to the Banach space dual, endowed with this matrix-normed structure, as the dual operator space. We shall write E ∗ for the dual operator space of E as well as for the Banach space dual of E, and hope it will be clear from the context which dual we are referring to. Proposition 14.8. Let E be an operator space, and let j: E → E ∗∗ be the canonical embedding j(e)(ϕ) = ϕ(e) for ϕ ∈ E ∗ . Then j is a complete isometry. Proof. If X = (xi j ) ∈ Mm,n (E), then ( j(xi j )) ∈ Mm,n (E ∗∗ ) is identiﬁed with the map Xˆ : E ∗ → Mm,n deﬁned by Xˆ (ϕ) = (ϕ(xi j )). Thus, we must prove Xˆ cb = (xi j ). Now if (ϕk, ) ∈ M pq (E ∗ ), then ( Xˆ (ϕk, )) = (ϕk (xi j )) ≤ cb (xi j ) where : E → M pq is given by (x) = (ϕk (x)). Since (ϕk ) = cb , we have that Xˆ cb ≤ (xi j ). Now to prove the other inequality we shall need Ruan’s theorem. So let γ : E → B(H) be a complete isometry. Pick subspaces Hq ⊆ Hq+1 with dim(Hq ) = q such that (xi j ) = (γ (xi j )) = supq (Pq γ (xi j )Pq ), where Pq : H → Hq denotes the projection. Choosing an orthonormal set {ek }∞ k=1 such that span{e1 , . . . , eq } = Hq , we obtain linear functionals γk (x) = γ (x)e , ek q q such that Pq γ (x)Pq ∼ = (γk, (x))k, =1 . Hence (γk, )k, =1 ∈ Mq (E ∗ ), (γk, ) ≤ γ cb = 1, and Xˆ cb ≥ sup ( Xˆ (γk, )) = sup (Pq γ (xi j )Pq ) = (xi j ). q

q

Blecher [19] proves the deeper fact that the image of the unit ball of Mm,n (E) is weak∗ -dense in the unit ball of Mm,n (E ∗∗ ) for all m and n. Not surprisingly, some of the most important operator spaces are those that as normed spaces are Hilbert spaces. We have already met two ways to make n-dimensional Hilbert space 2n into an operator space: MIN( 2n ) and MAX( 2n ). By Exercise 14.5, MIN( 2n )∗ = MAX( 2n ) and MAX( 2n )∗ = MIN( 2n ), so these are a dual pair. Another important dual pair are row and column Hilbert space. Consider the identiﬁcation of 2n with the operator space Mn,1 of the n × 1 matrices, i.e., column vectors of length n. A p × q matrix of vectors (vi j ) is normed by writing it as a pn × q scalar matrix under the identiﬁcation M p,q (Mn,1 ) = M pn,q . When the Hilbert space 2n is equipped with this operator space structure, we call it column Hilbert space, and it is generally denoted Cn instead of Mn,1 . Similarly, if we identify 2n with the operator space M1,n , i.e., row vectors of length n, then a p × q matrix of vectors (vi j ) is normed by writing it as a p × qn

198

Chapter 14. An Operator Space Bestiary

scalar matrix under the identiﬁcation M p,q (M1,n ) = M p,qn . The Hilbert space 2n equipped with this operator space structure is called row Hilbert space, and it is denoted Rn instead of M1,n . For a quick example of how these two structures differ, let {e1 , . . . , en } denote the canonical basis vectors for 2n , and consider the 1 × n matrix of vectors (e1 , . . . , en ). We have M1,n (Cn ) = Mn,n , and this is identiﬁed with the identity matrix, while M1,n (Rn ) = M1,n 2 and hence is identiﬁed with a row vector of length n 2 . Hence we have (e1 , . . . , en ) M1,n (Cn ) = 1, while (e1 , . . . , en ) M1,n (Rn ) = √ n, and thus the natural isometric identiﬁcation of Cn with Rn is not a complete isometry. In fact, if an n × n matrix A is regarded as a linear map from Cn to Rn , then Acb is equal to the Hilbert–Schmidt norm of A. For a proof of this fact, see [140]. We now wish to examine these two operator space structures from another viewpoint. Formally, Cn can be thought of as the image of 2n under the isometric map ϕ: 2n → Cn given by ϕ(ei ) = E i,1 . Given a matrix (vk ) ∈ M p,q ( 2n ), let Ai = (vk , ei ), 1 ≤ i ≤ n, denote the scalar p × q matrices obtained by taking the coefﬁcients of (vk ) with respect to the ith basis vector. The identiﬁcation n Ai ⊗ ei , and M p,q ( 2n ) = M p,q ⊗ 2n identiﬁes (vk ) = i=1 n n (ϕ(vk )) = id M p,q ⊗ ϕ Ai ⊗ ei = Ai ⊗ ϕ(ei ) i=1

=

n

Ai ⊗ E i1

i=1

in

M p,q ⊗ Mn,1 = M pn,q .

i=1

Thus (ϕ(vk )) is identiﬁed with the pn × q matrix   A1  ..   . , An and we have

∗ n n (vk )2M p,q (Cn ) = Ai ⊗ E i1 A j ⊗ E ji i=1 j=1 n n ∗ ∗ = Ai Ai ⊗ E 11 = Ai Ai . i=1 i=1 n A similar analysis for Rn shows that if (vk ) = i=1 Ai ⊗ ei is in M p,q ⊗ 2n , n then its image in Rn is i=1 Ai ⊗ E 1i , which can be identiﬁed with the p × qn

Chapter 14. An Operator Space Bestiary

199

matrix (A1 , . . . , An ), and ∗ n n n (vk )2M p,q (Rn ) = Ai ⊗ E 1i A j ⊗ E1 j = Ai Ai∗ . i=1 j=1 i=1 Given v ∈ Rn , deﬁne γ (v): Cn → C, via γ (v)(w) = v · w, where v · w denotes the matrix product of the row vector v with the column vector w. This gives a map γ : Rn → Cn∗ , which is clearly an isometry. Proposition 14.9. The map γ : Rn → Cn∗ is a complete isometry. Proof. Let (vi j ) ∈ M p,q (Rn ) = M p,qn . Then = (γ (vi j )): Cn → M p,q is the map given by (w) = (vi j · w), and we must prove that φcb = (vi j ) M p,qn . First note that if we form the qn × q matrix   w 0 ... 0 . .    0 w . . ..  wˆ =  .. . . . . , . . 0 . 0 ... 0 w then w ˆ = w and (w) = (vi j ) · wˆ (w) M p,q ≤ (vi j ) M p,q (Rn ) · w. Now if (wk, ) ∈ Mr,s (Cn ), then  (vi j )  ((wk, )) = ((vi j ) · wˆ k, ) =  0

is the actual matrix product. Thus,

0 .. .

 

 wˆ 11 . . . wˆ 1s ..    .. · . . , (vi j ) wˆ r 1 . . . wˆ r s

and hence ((wk )) ≤ (vi j ) M p,q (Rn ) (wˆ k ) Mr,s (Mqn,q ) . A moment’s reﬂection and a canonical shufﬂe show that (wˆ k ) Mr,s (Mqn,q ) = (wk ) Mr,s (Cn ) . Hence, cb ≤ (vi j ) M p,q (Rn ) . To see that cb ≥ (vi j ) M p,q (Rn ) , consider (E 11 , . . . , E n1 ) ∈ M1,n (Cn ). Then (E 11 , . . . , E n1 ) = 1, and so cb ≥ ((E 11 ), . . . , (E n1 )) M p,qn = ((vi j · E 11 ), . . . , (vi j · E n1 )) M p,qn = (A1 , . . . , An ) M p,qn .

200

Chapter 14. An Operator Space Bestiary

But the scalar matrices A1 , . . . , An are easily seen to be the coefﬁcients of n Ai ⊗ ei , and so this last quantity is (vi j ) M p,q (Rn ) by the above (vi j ) = i=1 calculation. In a similar fashion the matrix product pairing deﬁnes a complete isometry of Cn onto Rn∗ . It is valuable to look at the above pairing from the tensor viewpoint. Keepn Ai ⊗ ei and ing the notation as above, if (w) = (vi j · w) with (vi j ) = i=1 n n (wk ) = i=1 Bi ⊗ ei , then ((wk )) = i=1 Ai ⊗ Bi where Ai ⊗ Bi is the Kronecker tensor product of matrices discussed in Chapter 3. One of the most important properties of a Hilbert space is that it is conjugate linearly isometrically isomorphic to its own dual. Thus, given a Hilbert space H, one would like an operator space structure on H such that the map from H to H∗ is a conjugate linear complete isometry. Pisier [176] proved that such an operator space structure exists and is unique. The resulting operator space is called the operator Hilbert space and is denoted OH. n Ai ⊗ ei is in If {e1 , . . . , en } denotes the standard basis for 2n and i=1 M p,q ( 2n ), then n n Ai ⊗ ei = Ai ⊗ A¯ i , 2 i=1 i=1 O n

M p2 ,q 2

where A¯ denotes the complex conjugate of the matrix A, and A ⊗ A¯ is the p 2 × q 2 Kronecker tensor encountered earlier. That is, A ⊗ A¯ is the matrix whose ((i 1 , i 2 ), ( j1 , j2 )) entry is ai1 , j1 a¯ i2 , j2 for 1 ≤ i 1 , i 2 ≤ p and 1 ≤ j1 , j2 ≤ q. Pisier’s proof rests on a pretty Cauchy–Schwarz-type inequality for Kronecker tensors due to Haagerup [108] namely, 1/2 1/2 n n n Ai ⊗ Bi ≤ Ai ⊗ A¯ i Bi ⊗ B¯ i . i=1 i=1 i=1 In addition to the ﬁve operator space structures that we have already discussed on Hilbert space, there are many more. Among these the most natural to study are the homogeneous operator space structures. An operator space X is called homogeneous provided every bounded linear map T : X → X is completely bounded and T cb = T . For example, MIN(X ) and MAX(X ) are homogeneous operator spaces for every normed space (Exercise 14.1). In addition, row, column, and Pisier’s operator Hilbert space are homogeneous operator spaces. There exist many “exotic” homogeneous operator space structures on Hilbert space. Zhang [248] exhibits, for each integer n, two homogeneous operator

Chapter 14. An Operator Space Bestiary

201

space structures on 2n , say H1 and H2 , such that if T : H1 → H2 is linear with rank(T ) ≤ n − 1, then T cb = T , but I cb = I , where I denotes the identity. In addition, every pair of (n − 1)-dimensional subspaces of H1 and H2 are completely isometrically isomorphic, yet H1 and H2 are not completely isometrically isomorphic. For each symmetrically normed ideal J in the sense of Gohberg and Krein [105], Mathes and the author [142] exhibit two homogeneous operator space structures on 2 say H1 and H2 , such that T : H1 → H2 is completely bounded if and only if T ∈ J and the completely bounded norm of T is equivalent to its norm in J . Thus, in particular, for 1 < p < +∞ there exist homogeneous operator space structures on 2 such that T : H1 → H2 is completely bounded if and only if p 1/ p , it belongs to the Schatten p-ideal C p with T cb equivalent to ( ∞ n=1 sn ) where {sn } is the sequence of singular values of T . A characterization of the set of symmetric norms that one can achieve exactly as the cb norm between two homogeneous operator spaces is not known. It is also not very well understood how nonhomogeneous an operator space norm can be made. That is, given an operator space X , how small a subset CB(X, X ) can be of B(X, X ) is not clear. The most deﬁnitive result in this direction is due to Oikhberg [154]. He exhibits an operator space structure on Hilbert space such that the completely bounded maps from the space back to itself are just the scalar multiples of the identity plus the compacts. On the other hand it is quite easy to show that, independent of the operator space structure on Hilbert space, every map that is the sum of a scalar multiple of the identity and a trace class operator is necessarily completely bounded. These results are in sharp contrast to the Banach space setting, where it is still unknown if any Banach space X exists such that every operator in B(X, X ) is a scalar multiple of the identity plus a compact. Hilbert C∗ -Modules We close this chapter by studying some objects that play an important role in the study of C ∗ -algebras. To this end, let A be a unital C ∗ -algebra and let V be a vector space that is also a left A-module. Thus, in particular, we assume a1 · (v1 + v2 ) = a1 · v1 + a1 · v2 , (a1 + a2 ) · v1 = a1 · v1 + a2 · v2 , (a1 a2 ) · v1 = a1 · (a2 · v1 ), and (λa1 ) · v1 = λ(a1 · v1 ), for a1 , a2 ∈ A and v1 , v2 ∈ V , and λ ∈ C. We also assume that 1 · v = v for every v in V. Sometimes modules satisfying this latter condition are called unitary A-modules, but we ﬁnd this terminology can be confusing.

202

Chapter 14. An Operator Space Bestiary

We call V a Hilbert C ∗ -module provided it is equipped with a map , : V × V → A satisfying: (i) (ii) (iii) (iv) (v)

v1 + v2 , v3 = v1 , v3 + v2 , v3 , av1 , v2 = av1 , v2 , v1 , av2 = v1 , v2 a ∗ , v1 , v2 = v2 , v1 ∗ , v1 , v1 ≥ 0 and v1 , v1 = 0 if and only if v1 = 0,

for every v1 , v2 , v3 ∈ V and a ∈ A. We call such a map an A-valued inner product. We note that we have made no attempt to present a minimal set of axioms. For example, (ii) and (iv) taken together imply (iii). Thus, a Hilbert C ∗ -module is like a Hilbert space, where the role of C is played by A. Perhaps the simplest example of a Hilbert C ∗ -module over A is to let V = {(a1 , . . . , an ): ai ∈ A} with module action given by a(a1 , . . . , an ) = (aa1 , . . . , aan ) and an A-valued inner product given by (a1 , . . . , an ), (b1 , . . . , bn ) = a1 b1∗ + · · · + an bn∗ . Note that if we identify V = M1,n (A), then V ∗ = Mn,1 (A) and for v1 = (a1 , . . . , an ) and v2 = (b1 , . . . , bn ) in V we have that  ∗ b1 . v1 , v2 = (a1 , . . . , an ) ·  ..  = v1 v2∗ . bn∗ Thus, we see that V is identiﬁed with an operator space M1,n (A) in such a way that the A-valued inner product becomes an actual product between V and V ∗ . It is a well-known result in the theory of Hilbert C ∗ -modules that such a representation is always possible. Thus, in particular, every Hilbert C ∗ -module has a “natural” operator space structure. Using Ruan’s theorem, it is possible to give an intrinsic description of this operator space structure on a Hilbert C ∗ -module. Clearly, if we want v1 , v2 = v1 v2∗ , then for any X = (vi j ) ∈ M p,q (V ) we should have that X X ∗ ∈ M p, p (A) is the element given by q q ∗ ∗ vi v j = vi , v j . XX = =1

=1

Thus, we need that this latter matrix should be a positive element of M p (A) and q X 2 = ( =1 vi , v j ). This hopefully motivates the following theorem.

Chapter 14. An Operator Space Bestiary

203

Theorem 14.10. Let V be a Hilbert C ∗ -module over A. Then for every p, q q and X = (vi j ) ∈ M p,q (V ) the matrix ( =1 vi , v j ) ∈ M p (A)+ . If we deﬁne q X p,q = ( =1 vi , v j )1/2 , then V endowed with this family of matrix norms is an operator space. q Proof. To verify that the matrix ( =1 vi , v j ) is a positive element of M p (A), by Exercise 3.18 it is sufﬁcient to prove that for every p-tuple q p (a1 , . . . , a p ) of elements of A, we have that i, j=1 =1 ai vi , v j a ∗j is a p positive element of A. Setting x = i=1 ai vi , we see that this latter sum is p =1 x , x , which is positive by property (v). To see that v = v, v1/2 is a norm on V , we note that λv = |λ|v is clear and establish the triangle inequality. To this end, let v, w ∈ V and set X = ( vw ) ∈ M2,1 (V ). Since

v, v v, w X X∗ = ∈ M2 (A)+ , w, v w, w by Exercise 3.2 (ii) we have that v + w, v + w = v, v + v, w + w, v + w, w 2 ≤ v, v1/2 + w, w1/2 and thus v + w ≤ v + w. To see that the above quantity deﬁnes a norm on M p,q (V ), simply note that the “natural” matrix product makes M p,q (V ) a left M p (A)-module. For X = (vi j ) q and Y = (wi j ) in M p,q (V ) set X, Y p,q = ( k=1 vik , w jk ) ∈ M p (A). Since, formally, X, Y p,q = X · Y ∗ , it is readily seen that ·, · p,q is an M p (A)-valued inner product on M p,q (V ). Thus, by the last paragraph, we know that setting X p,q = X, X p,q

1/2

1/2 q = vik , v jk k=1

M p (A)

deﬁnes a norm on M p,q (A). The fact that this system of norms satisﬁes Ruan’s axioms is straightforward. In particular, notice that the L ∞ condition follows from the fact that for X ∈ M p,q (V ), and Y ∈ Mr,s (V ) we have that X ⊕ Y p+r,q+s = X ⊕ Y, X ⊕ Y p+r,q+s 1/2 = X, X p,q ⊕ Y, Y r,s 1/2 = max{X , Y }.

204

Chapter 14. An Operator Space Bestiary Notes

The theory of operator spaces is currently a very exciting and rapidly expanding area. See [88] and [193] for two recent texts in this ﬁeld. The constant α was ﬁrst introduced and studied in [163], where many of the early estimates of its values were obtained. Some small improvements can be found in [164], and the key improvement on the lower bound for α( 2n ) appears in [193]. The geometry of ﬁnite-dimensional operator spaces now plays an important role in many questions in the general theory of C ∗ -algebras. One of the most exciting breakthroughs was its use by Junge and Pisier [127] to prove that the maximal and minimal C ∗ -tensor norms of B(H) with itself are different, thus settling an old conjecture.

Exercises 14.1 Let V be a normed space, X an operator space, and let ϕ: X → MIN(V ), ψ: MAX(V ) → X be linear maps. Prove that ϕ = ϕcb and ψ = ψcb . Deduce that MIN(V ) and MAX(V ) are homogeneous. 14.2 Let V be a normed space, X an operator space, and let ϕ: MIN(V ) → X. Prove that ϕcb ≤ α(V )ϕ. 14.3 (Zhang) Let Fm denote the free group on m generators ! (m) " u 1 , . . . , u (m) m . Prove that the maps ϕ: MAX( 1n ) → Cu∗ (Fn ) and ψ: MAX( 1n ) → Cu∗ (Fn−1 ) given by ϕ((λ1 , . . . , λn )) = λ1 u n1 + · · · + λn u nn and + λ2 u (n−1) + · · · + λn−1 u (n−1) ψ((λ1 , . . . , λn )) = λ1 u (n−1) 1 2 n−1 + λn I

14.4

14.5 14.6

14.7

are complete isometries. ∗ (n) [Hint: Prove that the map sending u i(n−1) → u (n) induces a ∗n ui ∗ ∗ isomorphism of Cu (Fn−1 ) into Cu (Fn ).] Prove that the maps ϕ: MIN( 1n ) → C(Tn ) and ψ: MIN( 1n ) → C(Tn−1 ) given by ϕ((λ1 , . . . , λn )) = λ1 z 1 + · · · + λn z n and ψ((λ1 , . . . , λn )) = λ1 z 1 + · · · + λn−1 z n−1 + λn are complete isometries. Prove Proposition 14.7. (Blecher) Let V be a normed space. Prove that MIN(V )∗ = MAX(V ∗ ) and MAX(V )∗ = MIN(V ∗ ), completely isometrically. Use these identiﬁcations to prove that α(V ) = α(V ∗ ). Recall that for H, K Hilbert spaces, B(H, K) is an operator space. The map ϕ: B(C, H) → H via ϕ(T ) = T (1) for T ∈ B(C, H) is an isometry.

Exercises

205

Prove that B(C, 2n ) is completely isometrically isomorphic to Cn . Show similarly, that B( 2n , C) is completely isometrically isomorphic to Rn . We set Hc = B(C, H) and Hr = B(H, C), and call these operator spaces Hilbert column space and Hilbert row space, respectively. n n λi E i1 ) = i=1 λi E 1i . Prove that 14.8 Let γ : Cn → Rn be the map γ ( i=1 √ −1 γ is an isometry and γ cb = γ cb = n. 14.9 Let V be a Hilbert C ∗ -module over A. Prove the following analogue of the Cauchy–Schwarz inequality: v, ww, v ≤ w, wv, v. 14.10 Prove the analogue of Theorem 14.6 with H ∞ (G) replaced by P(G). Prove that the inclusion of P(G) into H ∞ (G) induces a completely isometric isomorphism between the quotient spaces P0 (G)/P00 (G) and ∞ (G). the quotient space H0∞ (G)/H00

Chapter 15 Injective Envelopes

In this chapter we take a closer look at injectivity and introduce injective envelopes and C ∗ -envelopes of operator systems, operator algebras, and operator spaces. Loosely speaking, the injective envelope of an object is a “minimal” injective object that contains the original object. The C ∗ -envelope of an operator algebra is a generalization of the Silov boundary of a uniform algebra. The C ∗ -envelope of an operator algebra A is the “smallest” C ∗ -algebra that contains A as a subalgebra, up to completely isometric isomorphism. These ideas will be made precise in this chapter. Many of the ideas of this chapter are derived from the work of M. Hamana [112]. Injectivity is really a categorical concept. Suppose that we are given some category C consisting of objects and morphisms. Then an object I is called injective in C provided that for every pair of objects E ⊆ F and every morphism ϕ: E → I , there exists a morphism ψ: F → I that extends ϕ, i.e., such that ψ(e) = ϕ(e) for every e in E. If we let S denote the collection of operator systems and deﬁne the morphisms between operator systems to be the completely positive maps, then since the composition of completely positive maps is again completely positive, we shall have a category, which we call the category of operator systems. Arveson’s extension theorem (Theorem 7.5) for completely positive maps is equivalent to the statement that B(H) is injective in the category S. The earlier deﬁnition of injectivity for C ∗ -algebras that was given in Chapter 7 was really injectivity in the category S. Similarly, we let O denote the category whose objects consist of operator spaces and whose morphisms are the completely bounded maps. Wittstock’s extension theorem (Theorem 8.2) for completely bounded maps shows that B(H) is also injective in the category O. However, note that for it to be injective in O, we only need that every completely bounded map has a completely bounded extension, but not necessarily of the same cb norm! To capture the full force of 206

Chapter 15. Injective Envelopes

207

Wittstock’s extension theorem, we need a slightly different category. We let O1 denote the category whose objects are operator spaces and whose morphisms are the completely contractive maps. Now by a simple scaling argument, it is easy to see (Exercise 15.1) that I is injective in O1 if and only if every completely bounded map into I has a completely bounded extension of the same cb norm. Thus, Wittstock’s extension theorem is equivalent to the statement that B(H) is injective in O1 . In this chapter we focus on injectivity in O1 . For operator systems it might seem natural to also study injectivity in S or in the category S1 consisting of operator systems and unital completely positive maps, but it turns out that for operator systems injectivity in these three categories is equivalent (Proposition 15.1). We begin with an elementary characterization of injectivity for operator systems. Proposition 15.1. Let S ⊆ B(H) be an operator system. Then the following are equivalent: (i) (ii) (iii) (iv)

S is injective in O1 , S is injective in S, S is injective in S1 , there exists a completely positive projection φ: B(H) → S onto S.

Proof. We begin by proving that (i) is equivalent to (iv). Assuming (i), then the identity map from S to S extends to a completely contractive map φ: B(H) → S. Since φ extends the identity map, φ is a projection onto S. Since φ(1) = 1, φ must be completely positive. Conversely, assume (iv), and suppose we are given operator spaces E ⊆ F and a completely contractive map, γ : E → S. Then γ has a completely contractive extension ψ: F → B(H), and φ ◦ ψ: F → S is the desired completely contractive extension of γ into S. The proof of the equivalence of (ii) and (iii) to (iv) is similar. Thus, for operator systems we have that injectivity in S, in S1 , and in O1 are all equivalent. For this reason we shall continue to simply call such systems injective. However, it is important to remark that there is an example due to Huruya [125] of a C ∗ -algebra that he proves is not injective, but is easily seen [102] to be injective in O. Currently, there is little that is known about the C ∗ algebras that are injective in O. Every C ∗ -algebra that is completely boundedly linearly isomorphic to a C ∗ -algebra that is injective in O1 can be easily seen to

208

Chapter 15. Injective Envelopes

be injective in O, but it is not known if this characterizes the C ∗ -algebras that are injective in O. The following result shows that every injective operator system is in an appropriate sense a C ∗ -algebra. Theorem 15.2 (Choi–Effros). Let S ⊆ B(H) be an injective operator system, and let φ: B(H) → S be a completely positive projection onto S. Then setting a ◦ b = φ(a · b) deﬁnes a multiplication on S, and S together with this multiplication and its usual ∗-operation is a C ∗ -algebra. Moreover, the identity map from S to the C ∗ -algebra (S, ◦) is a unital complete order isomorphism. Proof. Clearly, for a, b in S, a ◦ b is in S. Distributivity is clear, and clearly a ◦ 1 = 1 ◦ a = a. To show that ◦ deﬁnes a multiplication on S it remains to show associativity, a ◦ (b ◦ c) = (a ◦ b) ◦ c, i.e., that φ(aφ(bc)) = φ(φ(ab)c). We claim for any x in B(H) and a in S that φ(φ(x)a) = φ(xa) and φ(aφ(x)) = φ(ax). Assuming the claim, we have that φ(aφ(bc)) = φ(abc) = φ(φ(ab)c), and thus associativity follows. To prove the claim recall the Schwarz inequality, ψ(y ∗ y) − ψ(y)∗ ψ(y) ≥ 0 for any unital completely positive map ψ. Applying this to ψ = φ (2) and y = ∗ ( a0 x0 ) yields φ(aa ∗ ) φ(ax) aφ(x) aa ∗ − ≥ 0. φ(x ∗ a ∗ ) φ(x ∗ x) φ(x)∗ a ∗ φ(x)∗ φ(x) Applying φ (2) to this inequality yields 0 φ(ax) − φ(aφ(x)) ≥ 0. φ(x ∗ a ∗ ) − φ(φ(x)∗ a ∗ ) φ(x ∗ x) − φ(φ(x)∗ φ(x)) The positivity of this matrix forces φ(ax) − φ(aφ(x)) = 0, and hence, since φ is self-adjoint, the claim follows. Next we verify the C ∗ condition, a ∗ ◦ a = a 2 . Clearly, a ∗ ◦ a = φ(a ∗ a) ≤ a ∗ a = a 2 . But, again by the Schwarz inequality, φ(a ∗ a) ≥ φ(a)∗ φ(a) = a ∗ a and hence a ∗ ◦ a = φ(a ∗ a) ≥ a ∗ a = a 2 , and the proof that (S, ◦) is a C ∗ -algebra is complete. Clearly the identity map from S to (S, ◦) is an isometry. We now argue that it is a complete isometry. To this end consider Mn (S) ⊆ Mn (B(H)) = B(H(n) ) and φ (n) : B(H(n) ) → Mn (S). By the above, Mn (S) is a C ∗ -algebra with product

Chapter 15. Injective Envelopes

209

A ◦n B = φ (n) (A · B), and this C ∗ -algebra is isometrically isomorphic to the operator system Mn (S). But for A = (ai j ), B = (bi j ) we have n n n (n) A ◦n B = φ aik bk j = φ(aik bk j ) = aik ◦ bk j . k=1

k=1

k=1

Thus, (Mn (S), ◦n ) is the C ∗ -algebra tensor product of Mn and (S, ◦), and hence, by the uniqueness of the C ∗ -norm on Mn ((S, ◦)), the identity map from S to (S, ◦) is an n-isometry for all n. Since the identity map is a unital complete isometry, it is a complete order isomorphism as claimed. We now turn our attention to operator spaces. Since in this chapter we are mainly concerned with injectivity in O1 , we shall also simply refer to an operator space as injective when it is injective in O1 . This is also consistent with the terminology used in most of the operator space literature. Arguing as in the proof of Proposition 15.1, it is easily seen that an operator space E ⊆ B(H) is injective if and only if there exists a completely contractive projection of B(H) onto E (Exercise 15.2). We wish to construct the injective envelope of an operator space. There are many equivalent ways to deﬁne these objects. We prefer to take a deﬁnition with the fewest hypotheses and deduce the other properties as consequences. Given an operator space F, we say that (E, κ) is an injective envelope of F provided that (i) E is injective in O1 , (ii) κ: F → E is a complete isometry, (iii) if E 1 is injective with κ(F) ⊆ E 1 ⊆ E, then E 1 = E. Any (E, κ) satisfying (ii) is called an extension of F. When (E, κ) satisﬁes (i) and (ii), then it is called an injective extension of F. Identifying F with κ(F), we often simply regard E as containing F, with the understanding that now κ is simply the inclusion of F into E. Thus, an injective envelope of F is, loosely, a minimal injective containing F. To prove the existence of such an object, one would like to simply invoke Zorn’s lemma. But to use this approach it would be necessary to prove that if {E λ } is a decreasing chain of injectives with F ⊆ E λ , then λ E λ is also injective, and this is not clear. The use of minimal F-seminorms, which we deﬁne below, is a way to ﬁnesse this problem. Assume that F ⊆ B(H). We call a map ϕ: B(H) → B(H) an F-map provided that ϕ is completely contractive and ϕ(x) = x for all x in F. An F-map ϕ such that ϕ ◦ ϕ = ϕ is called an F-projection. Thus, an F-projection ϕ is a

210

Chapter 15. Injective Envelopes

completely contractive projection onto E = ϕ(B(H )), with F ⊆ E, but we do not demand F = E. We deﬁne a partial order on F-projections by setting ψ ≺ ϕ provided that ψ ◦ ϕ = ψ = ϕ ◦ ψ. Given an F-map ϕ, we deﬁne a F-seminorm pϕ on B(H) by setting pϕ (x) = ϕ(x) . There is a natural partial order on seminorms, deﬁned by p ≤ q if and only if p(x) ≤ q(x) for all x. Proposition 15.3. Let F ⊆ B(H) be an operator space. Then there exist minimal F-seminorms on B(H). Proof. Let ϕλ : B(H) → B(H) be F-maps such that pϕλ is a decreasing chain of F-seminorms. Recall the BW topology of Chapter 5. Since C B1 (B(H), B(H)) is BW-compact, {ϕλ } has a subnet {ϕλµ } converging to, say, ϕ. Clearly, ϕ is an F-map, and since |ϕ(x)h, k| = limµ |ϕλµ (x)h, k| ≤ lim infµ ϕλµ (x) h k , it follows that pϕ ≤ pϕλ for all λ. Thus, every decreasing chain of Fseminorms has a lower bound, and it follows by Zorn’s lemma that minimal F-seminorms exist. Theorem 15.4. Let F ⊆ B(H) be an operator space. If ϕ: B(H) → B(H) is a F-map such that pϕ is a minimal F-seminorm, then ϕ is a minimal F-projection and the range ϕ(B(H)) of ϕ is an injective envelope of F. Proof. We begin by proving that ϕ is a F-projection. Since ϕ ◦ ϕ is also an F-map and ϕ(ϕ(x)) ≤ ϕ(x) , we must have that ϕ ◦ ϕ(x) = ϕ(x) for all x ∈ B(H). Set ϕ (k+1) = ϕ (k) ◦ ϕ, then ϕ (k) (x) = ϕ(x) for all k ≥ 1. Set ψn (x) = [ϕ(x) + · · · + ϕ (n) (x)]/n; then ψn (x) ≤ ϕ(x) , and so ψn (x) = ϕ(x) , too. Hence, ϕ(x) − ϕ ◦ ϕ(x) = ϕ(x − ϕ(x)) = ψn (x − ϕ(x)) ϕ(x) + · · · + ϕ (n) (x) ϕ (2) (x) + · · · + ϕ (n+1) (x) − = n n ≤

2 ϕ(x) → 0. n

Thus, ϕ(x) − ϕ ◦ ϕ(x) = 0, and it follows that ϕ is an F-projection. Now suppose that ψ is a F-projection with ψ ≺ ϕ, so that ψ ◦ ϕ = ψ = ϕ ◦ ψ. Since ψ(x) = ψ(ϕ(x)) ≤ ϕ(x) , we have that ψ(x) = ϕ(x) for all x. Finally, ϕ(x) − ψ(x) = ϕ(ϕ(x) − ψ(x)) = ψ(ϕ(x) − ψ(x)) = ψ(x) − ψ(x) = 0, and hence ϕ(x) = ψ(x) for all x. Thus, ϕ is a minimal F-projection.

Chapter 15. Injective Envelopes

211

Since B(H) is injective in O1 , and ϕ is a completely contractive projection, it follows readily that ϕ(B(H)) is injective in O1 as well. Now assume that F ⊆ E 1 ⊆ ϕ(B(H)) with E 1 injective in O1 . Then the identity map from E 1 to E 1 extends to a completely contractive projection γ from B(H) to E 1 . Since γ ◦ ϕ is a F-map and γ (ϕ(x)) ≤ ϕ(x) , we have that γ ◦ ϕ(x) = ϕ(x) by minimality of the seminorm pϕ . Since γ is an isometry on ϕ(B(H )) and since γ (ϕ(x) − γ ◦ ϕ(x)) = 0, we have ϕ(x) = γ (ϕ(x)), and it follows that E 1 = ϕ(B(H)). Hence, ϕ(B(H)) is an injective envelope of F. Note that we do not assert that if ϕ is a minimal F-projection then pϕ is a minimal F-seminorm. Now that we know the existence of an injective envelope, we turn our attention to its uniqueness and further properties. Lemma 15.5. Let F ⊆ B(H) be an operator space, with ϕ: B(H) → B(H) a F-map such that pϕ is a minimal F-seminorm. If γ : ϕ(B(H)) → ϕ(B(H)) is completely contractive and γ (x) = x for all x in F, then γ (ϕ(x)) = ϕ(x) for all x in B(H). Proof. Since γ (ϕ(x)) ≤ ϕ(x) , we have that γ (ϕ(x)) = ϕ(x) by minimality of pϕ . Hence γ is an isometry and, in particular, one-to-one. Since pγ ◦ϕ = pϕ is a minimal F-seminorm, Theorem 15.4 implies that γ ◦ ϕ is a projection. Hence γ ◦ ϕ = γ ◦ ϕ ◦ γ ◦ ϕ = γ ◦ γ ◦ ϕ, because ϕ is the identity on the range of γ ◦ ϕ. Thus, γ ◦ (ϕ − γ ◦ ϕ) = 0. But since γ is one-to-one, ϕ = γ ◦ ϕ and the result follows. Theorem 15.6. Let (E 1 , κ1 ) and (E 2 , κ2 ) be two injective envelopes of the operator space F. Then the map i: κ1 (F) → κ2 (F) given by i(κ1 (m)) = κ2 (m) extends uniquely to a completely isometric isomorphism of E 1 onto E 2 . Proof. Let F ⊆ B(H) with ϕ: B(H) → B(H) a F-map such that pϕ is a minimal F-seminorm. If we can prove that the map κ1 : F → E 1 has a unique extension to a completely isometric isomorphism γ1 : ϕ(B(H )) → E 1 , then the result will follow, since γ2−1 ◦ γ1 will then yield the desired map between E 1 and E 2 . By injectivity of E 1 , a completely contractive map γ1 : ϕ(B(H)) → E 1 extending κ1 exists, and since ϕ(B(H)) is injective, there is a completely contractive map β: E 1 → ϕ(B(H)) with β(κ1 (x)) = x for all x in F. Now β ◦ γ1 : ϕ(B(H)) → ϕ(B(H)) is completely contractive and ﬁxes, F so by Lemma 15.5 it is the identity on ϕ(B(H)).

212

Chapter 15. Injective Envelopes

Since β and γ1 are both completely contractive, it follows that γ1 must be a complete isometry. But the range of γ1 will then be an injective operator subspace of E 1 , and so, by the minimality of E 1 , γ1 must be onto. Hence, γ1 is a completely isometric isomorphism. The above result shows that the injective envelope really only depends on the operator space up to complete isometry. Thus, if F ⊆ B(H), κ: F → B(H1 ) is a complete isometry, and ψ: B(H1 ) → B(H1 ) is a κ(F)-map such that pψ is a minimal κ(F)-seminorm, then (ψ(B(H1 )), κ) is an injective envelope of F and so completely isometrically isomorphic to ϕ(B(H)) for any F-map ϕ with pϕ a minimal F-seminorm. Corollary 15.7 (Rigidity). Let (E, κ) be an injective envelope of F, and let ψ: E → E be completely contractive with ψ(κ(x)) = κ(x) for all x in F. Then ψ(e) = e for all e in E. Proof. By Lemma 15.5, this is true for ϕ(B(H)), but by Theorem 15.6, ϕ(B(H)) and E are completely isometrically isomorphic. Clearly, this property is pre served by completely isometric isomorphisms. This rigidity property can be used to characterize injective envelopes. Call an extension (E, κ) of F a rigid extension if whenever ϕ: E → E is completely contractive and ϕ(κ(x)) = κ(x) for all X in F, then the map ϕ is the identity on E. Rigid extensions are related to essential extensions. Let (E, κ) be an extension of F, let M be any operator space, and let ϕ: E → M be completely contractive. If ϕ ◦ κ: F → M a complete isometry implies that ϕ is a complete isometry, then we call (E, κ) an essential extension of F. Formally, these deﬁnitions of rigidity and essentiality both depend on the category O1 , but since we will not be studying these concepts in any other categories, we have omitted this dependence from the deﬁnition. Finally, we say that (E, κ) is the maximal extension of a space F with some property, if (E, κ) is an extension of F and whenever (M, β) is an extension of E and (M, β ◦ κ) has the property, then β(E) = M. Similarly, we call (E, κ) a minimal extension of a space F with some property, if (E, κ) is an extension of F and whenever κ(F) ⊂ M ⊂ E and M has the property, then M = E. Theorem 15.8. Let F be an operator space, and let (E, κ) be an extension of F. Then the following are equivalent:

Chapter 15. Injective Envelopes

213

(i) (E, κ) is an injective envelope of F, i.e., a minimal injective extension of F, (ii) (E, κ) is a maximal rigid extension of F, (iii) (E, κ) is a maximal essential extension of F, (iv) E is injective and (E, κ) is a rigid extension of F, (v) E is injective and (E, κ) is an essential extension of F. Proof. We only prove the equivalence of (i) and (iii); the remaining implications are left as Exercise 15.4. Assuming (i), let (M, β) be an extension of E, and assume that (M, β ◦ κ) is an essential extension of F. Since E is injective, the inverse of β extends to a completely contractive projection from M onto E. This map is a complete isometry on β ◦ κ(F) and so must be a complete isometry on M. Hence, β(E) = M. Conversely, assume that (E, κ) is a maximal essential extension of F, and let (M, β) be an injective envelope of F. By the proof that (i) implies (iii), we know that (M, β) is an essential extension of F. The map β ◦ κ −1 from κ(F) to β(F) must extend to a completely contractive map γ of E into M. But since E is an essential extension of F, γ must be a complete isometry on E. Thus (M, γ ) is an extension of E and an essential extension of F. Hence, by maximality, γ (E) = M, and so (E, κ) is an injective envelope of F. We let I (F) denote the (essentially unique) injective envelope in O1 of an operator space F. Some comments are now in order on injective envelopes of operator systems. If S is an operator system, then it is also an operator space. Thus, it makes sense to seek either a minimal injective in O1 , a minimal injective in S1 , or a minimal injective in S. To construct such an object in S, one could take 1 ∈ S ⊆ B(H) and consider completely positive S-projections and mimic the proofs of 15.3– 15.7. But since these maps are necessarily unital, they are completely contractive, and we are back to the situation of an operator space. Thus, we see that there is nothing to be gained by reproving theorems about injective envelopes of operator systems in S or S1 . We do record one observation. Proposition 15.9. Let S be an operator system. Then the injective envelope I (S) of S in O1 is an injective operator system and hence is completely orderisomorphic to a C ∗ -algebra. Proof. Let 1 ∈ S ⊆ B(H). Then I (S) = ϕ(B(H)) for some S-projection ϕ that is completely contractive and hence completely positive. Thus, ϕ(B(H)) is an

214

Chapter 15. Injective Envelopes

operator system, and hence I (S) is a C ∗ -algebra under the product a ◦ b = ϕ(ab) by Theorem 15.2. One consequence of the above result is that if one “forgets” the order structure on S and only remembers the operator space structure, and embeds S completely isometrically via ψ: S → B(H) with say ψ(1) = 1, then I (ψ(S)) is completely isometric to I (S) and so is still completely isometrically isomorphic to a C ∗ -algebra. The next few results record how injective envelopes respect algebraic and C ∗ -structures. Proposition 15.10. Let 1 ∈ A ⊆ S ⊆ B(H) with S an operator system and A a subalgebra of B(H). Then the inclusion of A into the C ∗ -algebra I (S) is a completely isometric ∗ -isomorphism. Proof. Let ϕ: B(H) → B(H) be a S-projection such that ϕ(B(H)) is a copy of I (S). Then the multiplication on I (S) is given by x ◦ y = ϕ(x y). Hence for a, b in A, a ◦ b = ϕ(a · b) = a · b. In the case of a unital C ∗ -algebra A, if we set S = A, then the above result implies that A is a C ∗ -subalgebra of I (A). The next result explains the case of nonunital C ∗ -algebras. Proposition 15.11. Let A ⊆ B(H) be a nonunital C ∗ -algebra such that AH is dense in H, and let A1 denote the algebra obtained by adjoining the unit to A. If ϕ: B(H) → B(H) is an A-map, then ϕ is an A1 -map. Consequently, the inclusion of A into A1 extends uniquely to deﬁne a complete isometry of I (A) onto I (A1 ). Thus, I (A) is a unital C ∗ -algebra in the ◦-product with A1 a C ∗ -subalgebra of I (A). Proof. If every A-map is an A1 -map, then any minimal A-seminorm is a minimal A1 -seminorm and the last claim follows from the characterization of the injective envelope in terms of minimal seminorms. Thus, it is enough to prove that if ϕ is an A-map, then ϕ(I ) = I . Let {eα } be a contractive, positive approximate identity for A. Since AH is dense in H, we have eα → I strongly. Since {eαn } is also a contractive approximate identity for A, we have eαn → I strongly for all n.

Chapter 15. Injective Envelopes

215

By considering the power series for e x , we see that for t real ϕ(eiteα ) = ϕ(I ) +

∞ (iteα )n n=1

n!

→ ϕ(I ) + (eit − 1)I

in the strong operator topology. Since eiteα is unitary and ϕ is contractive, we have that (ϕ(I ) − I ) + eit I ≤ 1 for all t. Now it is easy to check that X + eit I ≤ 1 for all t implies that X = 0. Hence ϕ(I ) = I and the proof is complete. Next we examine how the injective envelope respects module actions. First suppose that 1 ∈ A ⊆ S ⊆ B(H) with A a C ∗ -subalgebra of B(H) and S a concrete operator system such that A · S ⊆ S. In this case, we also have that S · A = S ∗ · A∗ = (AS)∗ ⊆ S ∗ = S, so that S is an A-bimodule. Clearly, for any (ai j ) ∈ Mn (A) and (si j ) ∈ Mn (S)+ we have that (ai j ) · (si j ) · (ai j )∗ ∈ Mn (S)+ . We wish to characterize such situations abstractly. Assume we are given an abstract operator system S, with unit denoted by e, and a unital C ∗ -algebra A such that S is an A-bimodule. We denote the bimodule action and multiplication by ·, so that (a1 · a2 ) · s = a1 · (a2 · s), and we assume 1 · s = s. We call such an S an operator A-system provided that, in addition, a · e = e · a and for all n and for any (ai j ) ∈ Mn (A), (si j ) ∈ Mn (S)+ we have that n ∗ ∗ aik · sk · a j ∈ Mn (S)+ . (ai j ) · (si j ) · (ai j ) = k, =1

Theorem 15.12. Let A be a unital C ∗ -algebra, let S be an operator Asystem, and let (I (S), ◦) denote the injective envelope of S, with ◦ denoting the multiplication that makes I (S) into a C ∗ -algebra. Then the map π: (A, ·) → (I (S), ◦), π (a) = a · e, is a ∗-homomorphism. Proof. We must show that (a1 a2 ) · e = (a1 · e) ◦ (a2 · e). Because the unitaries span A, it will be enough to verify the above identity when a1 is unitary. Fix a unitary u ∈ A, and deﬁne : M2 (S) → M2 (S) via ∗ s1 s2 u · s1 · u ∗ u · s2 u 0 u 0 s 1 s2 . = = s3 s4 s3 · u ∗ s4 s3 s4 0 1 0 1 Since S is an operator A-system, is completely positive. Thus, we may extend to a completely positive map on I (M2 (S)) = M2 (I (S)), which we still denote

216

Chapter 15. Injective Envelopes

by . Since ﬁxes C⊕C=

λe 0 : λ, µ ∈ C , 0 µe

it is a bimodule map over this C ∗ -algebra by Corollary 3.20. Using the fact that is a bimodule map, or arguing as in the proof of Theorem 8.3, one ﬁnds that there exist maps ϕi : I (S) → I (S), i = 1, 2, 3, 4, such that ϕ1 (x1 ) ϕ2 (x2 ) x1 x2 = . x3 x4 ϕ3 (x3 ) ϕ4 (x4 ) Clearly, ϕ1 and ϕ4 are completely positive. Since ϕ4 (s4 ) = s4 for all s4 in S, by rigidity, ϕ4 = id I (S) . Thus, ﬁxes λe 0 C ⊕ I (S) = : λ ∈ C, x ∈ I (S) 0 x and so it is a bimodule map over this C ∗ -algebra. Hence, 0 ϕ2 (x) 0 x 0 e 0 0 = = ◦ 0 0 0 0 0 0 0 x 0 e 0 0 0 ϕ2 (e) ◦ x . = ◦ = 0 0 0 0 0 x Setting x = a2 · e, we have that (u · a2 ) · e = u · (a2 · e) = ϕ2 (a2 · e) = ϕ2 (e) ◦ (a2 · e) = (u · e) ◦ (a2 · e), and the proof is complete.

Corollary 15.13. Let A be a unital C ∗ -algebra, and let S be an operator A-system with unit e. Then there exists a Hilbert space H, a complete order isomorphism γ : S → B(H) with γ (e) = I H , and a unital ∗-homomorphism π: A → B(H) such that π(a)γ (s) = γ (a · s). Proof. Represent the C ∗ -algebra (I (S), ◦) ∗-isomorphically on a Hilbert space, and apply Theorem 15.12. We call (π, γ ) a representation of the operator A-system. The next representation theorem concerns operator space bimodules. Suppose that A ⊆ B(H), B ⊆ B(K) are unital C ∗ -subalgebras and X ⊆ B(K, H) is a subspace. If A · X · B ⊆ X , then X is an A-B-bimodule. Given (ai j ) ∈

Chapter 15. Injective Envelopes

217

Mn (A), (xi j ) ∈ Mn (X ), and (bi j ) ∈ Mn (B), we have that (ai j ) · (xi j ) · (bi j ) ≤ (ai j ) (xi j ) (bi j ) , where

(ai j ) · (xi j ) · (bi j ) =

n

aik xk b j .

k, =1

Such a module action is called completely contractive. We wish to characterize this situation abstractly. So let A, B be unital C ∗ -algebras, and let X be an operator space that is also an A-B-bimodule. We call X an operator A-B-bimodule provided that the module action is completely contractive. Theorem 15.14 (Christensen–Effros–Sinclair representation theorem). Let A, B be unital C ∗ -algebras, and let X be an operator A-B-bimodule. Then there exist Hilbert spaces H and K, unital ∗-isomorphisms π : A → B(H), ρ: B → B(K), and a complete isometry γ : X → B(K, H) such that γ (a · x · b) = π (a)γ (x)ρ(b). Proof. We imitate the proof of Ruan’s theorem with A and B playing the role of the scalars. To this end let a x S= : a ∈ A, b ∈ B, x, y ∈ X . y∗ b Setting b · y ∗ · a = (a ∗ · y · b∗ )∗ makes X ∗ an operator B-A-bimodule. We wish to make S into a matrix ordered ∗-vector space with Archimedean order unit 1A 0 . e= 0 1B Let

a x y∗ b

∗

a∗ y = , x ∗ b∗

and deﬁne Cn ⊆ Mn (S)h by saying (ai j ) (xi j ) ∈ Cn (yi∗j ) (bi j ) if and only if (yi∗j ) = (xi j )∗ and ((ai j ) + ε In )−1/2 (xi j )((bi j ) + ε In )−1/2 ≤ 1 for all > 0.

218

Chapter 15. Injective Envelopes

If one copies the proof of Ruan’s theorem mutatis mutandis, then one ﬁnds that all the axioms for an abstract operator system are satisﬁed by S (Exercise 15.11). The identiﬁcation of A ⊕ B with {( a0 b0 ): a ∈ A, b ∈ B} makes S an operator A ⊕ B-system. That is, deﬁne a x a1 aa2 a1 xb2 . · (a2 ⊕ b2 ) = (a1 ⊕ b1 ) · ∗ b1 y ∗ a2 b1 bb2 y b Now take any representation (θ, γ ) of S on a Hilbert space. The ∗-homomorphism θ of A ⊕ B decomposes the Hilbert space as H ⊕ K so that π(a) 0 θ (a ⊕ b) = , 0 ρ(b) where π: A → B(H) and ρ: B → B(K). Since 0 x 1 0 0 x 0 0 γ =γ · · 0 0 0 0 0 0 0 1 π(1) 0 0 x 0 0 = γ , 0 0 0 0 0 ρ(1) we see that 0 x 0 γ1 (x) γ = 0 0 0 0

where γ1 : X → B(K, H).

It is readily checked that since γ is a complete order isomorphism, γ1 is a complete isometry, and that π(a)γ1 (x)ρ(b) = γ1 (axb). Note that π and ρ are one-to-one, since A ⊕ B is a subset of S. We call (π, ρ, γ ) a representation of the operator A-B-bimodule X . We turn our attention now to Hamana’s construction of the C ∗ -envelope of an operator algebra. To motivate this material we begin with the example of the disk algebra, A(D). Clearly, A(D) can be regarded, up to completely isometric isomorphism, as either the uniformly closed subalgebra of C(D− ), or that of C(T), generated by the coordinate function z. In both cases A(D) separate the points on the underlying spaces, and so the C ∗ -algebra generated by A(D) will be C(D− ) in the ﬁrst case and C(T) in the second case. Thus, the C ∗ -algebra generated by an operator algebra can depend on the representation. Restricting a function in C(D− ) to ∂D− = T deﬁnes a ∗-homomorphism from C(D− ) onto C(T), which is the identity map on A(D). Thus, C(T) is a “smaller” representation of A(D) in the sense that it is a C ∗ -quotient of C(D− ).

Chapter 15. Injective Envelopes

219

We shall prove that among all the possible C ∗ -algebras that completely isometric representations of an operator algebra can generate, there exists a “smallest” such C ∗ -algebra in the sense that it is a universal quotient. To be more precise, given an operator algebra A, we shall prove that there exists a completely isometric representation γ : A → B(H) such that if ρ: A → B(K) is any other completely isometric representation, then there exists an onto ∗-homomorphism π : C ∗ (ρ(A)) → C ∗ (γ (A)) with π (ρ(a)) = γ (a). It is easy to see that if such a representation γ exists, then C ∗ (γ (A)) is uniquely determined up to ∗-isomorphism. That is, if γ1 and γ2 were two representations with the above property, then there would exist onto ∗-homomorphisms π1 : C ∗ (γ1 (A)) → C ∗ (γ2 (A)) with π1 (γ1 (a)) = γ2 (a) and π2 : C ∗ (γ2 (A)) → C ∗ (γ1 (A)) with π2 (γ2 (a)) = γ1 (a), but these two equations force π2 ◦ π1 = idC ∗ (γ1 (A)) and π1 ◦ π2 = idC ∗ (γ2 (A)) , and so π1 and π2 are ∗-isomorphisms. Proposition 15.15. Let 1 ∈ A ⊆ B(H) be an operator algebra. Then I (A) = I (A + A∗ ) is a C ∗ -algebra, and the inclusion of A into the C ∗ -algebra I (A) is a completely isometric isomorphism. Proof. Since 1 ∈ A, any completely contractive ϕ: B(H) → B(H) that ﬁxes A is completely positive, and hence for a and b in A ϕ(a + b∗ ) = ϕ(a) + ϕ(b)∗ = a + b∗ . Thus, every A-map is actually an (A + A∗ )-map, and so I (A) = I (A + A∗ ). The fact that the inclusion is a completely isometric isomorphism follows from Proposition 15.10. It is worth noticing that if ρ: A → B(H) is a completely isometric isomorphism, then the operator systems A + A∗ and ρ(A) + ρ(A)∗ are completely order-isomorphic via the map a + b∗ → ρ(a) + ρ(b)∗ . Thus, the operator system A + A∗ is really well deﬁned independent of the particular representation of A. This is not the case for the operator system A∗ A = span{a ∗ b: a, b ∈ A}, as can be seen by regarding A(D) ⊆ C(T) versus A(D) ⊆ C(D− ). Given an operator algebra A, we let Ce∗ (A) denote the C ∗ -subalgebra of I (A) generated by A. We call Ce∗ (A) the C ∗ -envelope of A. Theorem 15.16 (Hamana’s theorem). Let A be an operator algebra with unit, let ρ: A → B(H) be a completely isometric homomorphism, and let C ∗ (ρ(A)) denote the C ∗ -subalgebra of B(H) generated by ρ(A). Then there exists a

220

Chapter 15. Injective Envelopes

unique onto ∗-homomorphism π : C ∗ (ρ(A)) → Ce∗ (A) with π(ρ(a)) = a for all a in A. Proof. Once we show the existence of π, the fact that it is unique and onto follows from the fact that the C ∗ -algebras are generated by ρ(A) and A, respectively. By injectivity of I (A), there is a unital completely contractive (and hence completely positive) map ϕ: B(H) → I (A) with ϕ(ρ(a)) = a for a in A. Similarly, there is a completely positive map ψ: I (A) → B(H) with ψ(a) = ρ(a) for a in A. Since ϕ ◦ ψ(a) = a for a in A, by rigidity ϕ ◦ ψ = id I (A) . We claim that ϕ(ρ(a)∗ ρ(a)) = a ∗ a and ϕ(ρ(a)ρ(a)∗ ) = aa ∗ . Indeed, by the Schwarz inequality, ϕ(ρ(a)∗ ρ(a)) ≥ ϕ(ρ(a))∗ ϕ(ρ(a)) = a ∗ a, while ψ(a ∗ a) ≥ ψ(a)∗ ψ(a) = ρ(a)∗ ρ(a). Applying ϕ to the second inequality yields a ∗ a = ϕ ◦ ψ(a ∗ a) ≥ ϕ(ρ(a)∗ ρ(a)) ≥ a ∗ a and so ϕ(ρ(a)∗ ρ(a)) = a ∗ a. Similarly, ϕ(ρ(a)ρ(a)∗ ) = aa ∗ . Recalling Theorem 3.18 (iii), on the multiplicative domains of completely positive maps, these equalities imply that ϕ is a ∗-homomorphism when re stricted to C ∗ (ρ(A)). Thus, by the above theorem we see that Ce∗ (A) is a universal quotient for the C -algebra generated by any completely isometric representation of A. Some examples are helpful. We begin by showing how the classic theory of the Silov boundary of a uniform algebra follows from Hamana’s theorem. Recall that if X is a compact Hausdorff space, then a closed subalgebra, A ⊆ C(X ) is called a uniform algebra on X provided that 1 ∈ A and A separates points, i.e., for every x = y there exists f ∈ A with f (x) = f (y). ∗

Corollary 15.17 (Silov). Let A ⊆ C(X ) be a uniform algebra on X . Then there exists a unique compact subset Y ⊆ X such that: (i) the restriction homomorphism r : C(X ) → C(Y ), r ( f ) = f |Y is isometric on A, (ii) if W ⊆ X is any other set such that (i) holds, then Y ⊆ W . Moreover, if ρ: A → C(X 1 ) is any isometric isomorphism such that ρ(A) is a uniform algebra on X 1 , and Y1 ⊆ X 1 is the corresponding set for ρ(A), then there exists a homeomorphism h: Y1 → Y such that ρ( f )|Y1 = f ◦ h. Proof. Since the range of ρ is a commutative C ∗ -algebra, any isometric map is automatically completely isometric.

Chapter 15. Injective Envelopes

221

By Hamana’s theorem, there exists a ∗-homomorphism π : C(X ) → Ce∗ (A), onto. Hence, Ce∗ (A) ∼ = C(Z ) for some compact Hausdorff space Z . By the characterization of ∗-homomorphisms between commutative C ∗ -algebras (see [76]), there exists a one-to-one continuous function h: Z → X such that π (g) = g ◦ h. Let Y = h(Z ), and we have (i). Now suppose W satisﬁes (ii). Regarding A ⊆ C(W ), via the restriction map, we obtain π1 : C(W ) → C(Z ) and h 1 : Z → W such that π1 (g) = g ◦ h 1 . Let r : C(X ) → C(W ) denote the restriction map, so that r (g) = g ◦ i, where i: W → X denotes the inclusion. Since π = π1 ◦ r , it follows that h = i ◦ h 1 and so Y = h(Z ) = i(h 1 (Z )) must be a subset of W . The remainder of the proof follows similar arguments. The set Y , which only depends on A up to homeomorphism, is called the Silov boundary of A and denoted ∂ S A. It can be easily shown that ∂ S A(D) = T (Exercise 15.8). Corollary 15.18. Let A be a uniform algebra. Then Ce∗ (A) = C(∂ S A). By the above result we see that any time we represent a uniform algebra completely isometrically as an algebra of operators on a Hilbert space, the C ∗ -algebra generated by the image has a commutative C ∗ -algebra as a quotient. The following is another case where it is easy to recognize Ce∗ (A). Recall that a C ∗ -algebra is simple if it contains no nontrivial two-sided ideals. Proposition 15.19. Let A ⊆ B(H) be an operator algebra. If C ∗ (A) is simple, then Ce∗ (A) = C ∗ (A). Proof. There exists a ∗-homomorphism from C ∗ (A) onto Ce∗ (A), but because C ∗ (A) is simple, this map must be one-to-one. Thus, for example if A ⊆ Mn is the algebra of upper triangular matrices, then Ce∗ (A) = Mn . The existence of the C ∗ -envelope of an operator algebra with the properties of Hamana’s theorem was conjectured by Arveson [6], and he attempted to prove its existence by developing the theory of what are called boundary representations. Surprisingly, Hamana’s theorem does not guarantee the existence of boundary representations, and so many questions remain unanswered. We discuss these issues now. Assume A is an operator algebra and B = C ∗ (A). Any irreducible representation of Ce∗ (A) on a Hilbert space would give rise to an irreducible representation

222

Chapter 15. Injective Envelopes

of B. Since to determine the structure of Ce∗ (A) it is enough to know all its irreducible representations, we would like a condition that guarantees an irreducible representation of B corresponds to an irreducible representation of Ce∗ (A). We call an irreducible representation π: B → B(H) a boundary representation provided that if : B → B(H) is a completely positive map such that (a) = π (a) for all a in A, then π = . Assume that π is a boundary representation; then π |A will have a completely positive extension γ : Ce∗ (A) → B(H). Let ρ: B → Ce∗ (A) be the onto ∗-homomorphism, so that γ ◦ ρ: B → B(H) is a completely positive map that extends π |A . Hence, γ ◦ ρ = π , and it follows that γ must be a homomorphism and also be unique. Thus, every boundary representation of B gives rise to an irreducible representation of Ce∗ (A) that is also a boundary representation. Thus, one would hope that there exist enough boundary representations so that if one took the direct sum over all boundary representations, then that would yield a faithful, one-to-one representation of Ce∗ (A). Whether or not sufﬁciently many boundary representations exist is still unknown. To illustrate the difﬁculties, we consider the compact Hausdorff space X = D− × [0, 1] and the uniform algebra A on X given by A = { f ∈ C(X ): f (z, 0) ∈ A(D)}. It can be shown that X = ∂ S A, so that Ce∗ (A) = C(X ). Since Ce∗ (A) is commutative, the irreducible representations are all one-dimensional and are given by πx (g) = g(x) for points x in X . The boundary representations can be shown to correspond to evaluation at points of the form (z, t), t > 0, and (z, 0), |z| = 1, but none of the points (z, 0), |z| < 1, correspond to boundary representations. To see this last fact recall the Poisson kernel for |z| < 1, Pz (θ) =

1 1 − 1. + −iθ 1−e z 1 − e+iθ z¯

We have that Pz (θ) ≥ 0 and for f ∈ A(D),

2π 1 f (z) = f (eiθ )Pz (θ ) dθ. 2π 0 2π 1 iθ Thus, the map z : C(X ) → C deﬁned by z (g) = 2π 0 g(e , 0)Pz (θ) dθ is a completely positive map such that z ( f ) = πz ( f ) = f (z, 0) for f ∈ A, but z = π(z,0) . Thus, π(z,0) , |z| < 1, is not a boundary representation. For a general uniform algebra A, Arveson’s boundary representations can be shown to correspond to evaluations at what are called peak points, and these points are known to form a dense subset of ∂ S A called the Choquet boundary

Notes

223

(see [103] for deﬁnitions and results). Thus, a successful theory of boundary representations would lead to some noncommutative analogues of the fact that peak points are dense in the Silov boundary.

Notes The theory of the injective envelope of an operator system, in the category S, was developed in the work of M. Hamana [112]. Hamana not only recognized their importance, but used them to prove the existence of the C ∗ -envelope. The analogue of Theorem 15.8 for operator systems also appears in that work. These ideas were subsequently extended to the setting of operator spaces by Z.J. Ruan [204]. We will take a deeper look at these topics in Chapter 16. The original proof of the Christensen–Effros–Sinclair representation theorem used a representation theorem for multilinear completely bounded maps, which we will examine in Chapter 17. The proof that we present here seems to be new. There is still much to be understood about injectivity in the category O. For an operator space E ⊂ B(H) to be injective in this category it is necessary and sufﬁcient that there exist a completely bounded projection of B(H) onto E (compare with Exercise 15.2). As mentioned earlier, there exist C ∗ -algebras that are not injective in the usual sense (i.e., in O1 ) but are injective in O. However, it is known that for von Neumann algebras injectivity in O1 and injectivity in O are equivalent [61, 182]. A natural question that still remains open is whether or not every C ∗ -algebra that is injective in O is completely boundedly isomorphic to a C ∗ -algebra that is injective in O1 . All the known examples of C ∗ -algebras that are injective in O have this property. M. Frank and the author [102] pursue the study of injectivity in the category whose objects are left operator A-modules and whose morphisms are the completely bounded left A-module maps for a ﬁxed C ∗ -algebra A. In contrast to the above results, if A ⊂ B(H) and there exists a completely bounded projection of B(H) onto A that is also a left A-module map, then A is injective in the usual sense. It is shown that the usual injective envelope, I (A), is still the injective envelope in this new category. The authors also show that the theory of multipliers follows from properties of the injective envelope. In particular, if J is any essential ideal in A, then every multiplier of J is given as multiplication by a unique element of I (A). See [173] for the deﬁnitions of multipliers and essential ideals. M. Hamana [113], C. Zhang [247, 249], and D.P. Blecher [23] have developed the analogue of Hamana’s C ∗ -envelope of an operator algebra for operator spaces.

224

Chapter 15. Injective Envelopes Exercises

15.1 Verify that an operator space I is injective in O1 if and only if every completely bounded map into I has a completely bounded extension of the same completely bounded norm. 15.2 Let E ⊆ B(H) be an operator space. Prove that E is injective if and only if there exists a completely contractive map ϕ: B(H) → E such that ϕ(e) = e for all e ∈ E. 15.3 Let M ⊆ B(H) be an operator space, and let ϕ: B(H) → B(H) be an M-projection with pϕ a minimal M-seminorm. Assume that u ∈ B(H) is a unitary that commutes with M. Prove that if γ (x) = u ∗ ϕ(x)u, then pγ is a minimal M-seminorm and γ is a projection onto u ∗ ϕ(B(H))u. Prove that if ψ(x) = ϕ(u ∗ xu), then pψ is a minimal M-seminorm and ψ is a (possibly different) projection onto ϕ(B(H)). 15.4 Complete the proof of Theorem 15.8. 15.5 Let ϕ: B(H) → B(H) be a unital completely positive map that ﬁxes the compacts. Prove that ϕ is necessarily the identity map. [Hint: Let 0 ≤ P, and set P = {0 ≤ K ≤ P: K compact}. Prove that if K ≤ Q for all K ∈ P, then P ≤ Q.] 15.6 Let S ⊆ B(H) be an operator system that contains the compacts. Prove that I (S) = B(H). 15.7 Let A ⊆ B( 2 ) be the algebra of upper triangular operators. Prove that Ce∗ (A) = B( 2 ). 15.8 Prove that ∂ S A(D) = T. 15.9 Prove that if A ⊆ Mn then Ce∗ (A) is ﬁnite-dimensional. 15.10 Let A and B be unital C ∗ -algebras, and let ϕ: A → B be a complete order isomorphism with ϕ(1) = 1. Prove that there exists a ∗-homomorphism π: C ∗ (ϕ(A)) → A with π(ϕ(a)) = a for all a in A. 15.11 Verify that the set S deﬁned in the proof of Theorem 15.14 is a matrixordered *-vector space with complete Archimedean order unit e. 15.12 Let {u 1 , . . . , u n } denote the unitaries that generate Cu∗ (Fn ), and let M denote the (n + 1)-dimensional subspace generated by these unitaries and the identity I . If A ⊆ M2 (Cu∗ (Fn )) denotes the (n + 3)-dimensional operator algebra λI x A= : λ ∈ C, µ ∈ C, x ∈ M , 0 µI then prove that Ce∗ (A) = M2 (Cu∗ (Fn )). Thus, a ﬁnite-dimensional operator algebra can have an inﬁnite-dimensional C ∗ -envelope.

Chapter 16 Abstract Operator Algebras

In previous chapters we have given abstract characterizations of operator systems and operator spaces. In this chapter we wish to give an abstract characterization of operator algebras. That is, given an algebra A and norms on Mm,n (A), we desire necessary and sufﬁcient conditions for there to exist a Hilbert space H and a completely isometric isomorphism π: A → B(H). We present a theorem of Blecher, Ruan, and Sinclair (BRS) [31] that gives such necessary and sufﬁcient conditions. Thus, the BRS theorem plays exactly the same role for operator algebras that the celebrated Gelfand–Naimark–Segal theorem plays for C ∗ -algebras. Moreover, like those in the GNS theorem, the conditions in the BRS theorem are natural and quite often easy to verify. Brieﬂy, the BRS theorem says that the matrix norms on A must satisfy Ruan’s axioms for operator spaces (otherwise there wouldn’t even exist a linear complete isometry) and, for all n, the induced multiplication on Mn (A) must be contractive. That is, for A = (ai j ), B = (bi j ) in Mn (A) one has n aik bk j ≤ An Bn . k=1 n

This last condition is often summarized by saying that the multiplication on A is completely contractive. It is important not to confuse these results with other work on algebras of operators, where no matrix norms are assumed. In particular, Varopoulos [237] gives necessary and sufﬁcient conditions for a Banach algebra to have an isometric representation as an algebra of operators on a Hilbert space. This result has not been as widely used, perhaps because of the difﬁculty of verifying the conditions.

225

226

Chapter 16. Abstract Operator Algebras

There is still one outstanding problem in the isometric theory of operator algebras. If a unital Banach algebra B has the property that there exists a Hilbert space H and a unital isometric isomorphism π: B → B(H), then by von Neumann’s inequality, for every b ∈ B with b ≤ 1, and for every polynomial p we have p(b) ≤ p∞ , where the supremum is over the unit disk. A Banach algebra with this latter property is said to satisfy von Neumann’s inequality. Thus, Banach algebras that are isometrically isomorphic to an algebra of operators on a Hilbert space must satisfy von Neumann’s inequality. The problem is whether or not the converse is true. That is, if B is a Banach algebra that satisﬁes von Neumann’s inequality, then is B isometrically isomorphic to an algebra of operators on a Hilbert space? See [74] and [168] for two recent contributions to this problem. The key to the proof of the BRS theorem that we present here is a better understanding of the injective envelope of the “canonical” operator system associated to an operator space. Recall that if X ⊆ B(K, H) is a (concrete) operator space, then we may form the (concrete) operator system, in B(H ⊕ K), λIH x : λ, µ ∈ C, x, y ∈ X . SX = y ∗ µIK Given a complete isometry ϕ: X → B(K1 , H1 ), the operator system

λIH1 ϕ(x) Sϕ(X ) = : λ, µ ∈ C, x, y ∈ X ϕ(y)∗ µIK1 is completely order-isomorphic to S X via the map : S X → Sϕ(X ) deﬁned by λIH1 ϕ(x) λIH x = . y ∗ µIK ϕ(y)∗ µIK1 One way to see that is a complete order isomorphism is to recall that by Lemma 7.1, since ϕ and ϕ −1 are completely contractive, and −1 are completely positive. Thus, the operator system S X only depends on the operator space structure of X and not on any particular representation of X . This can also be seen in the proof of Ruan’s theorem, where for an arbitrary abstract operator space X we constructed this same operator system, abstractly, with only reference to the operator space structure of X . Since

0 λI H ∼ : λ, µ ∈ C C⊕C= 0 µIK

Chapter 16. Abstract Operator Algebras

227

is a C ∗ -subalgebra of S X , by Proposition 15.10 C ⊕ C will still be a C ∗ subalgebra of the C ∗ -algebra I (S X ), with IH 0 0 0 and 0 IK 0 0 corresponding to projections e1 and e2 , respectively, in the C ∗ -algebra I (S X ). We have that e1 + e2 is equal to the identity and e1 · e2 = 0. A few words on such a situation are in order. Let A be any unital C ∗ algebra with projections e1 and e2 satisfying e1 + e2 = 1, e1 · e2 = 0, and let π : A → B(H) be a one-to-one unital ∗-homomorphism. Setting H1 = π(e1 )H, H2 = π(e2 )H, we have that H = H1 ⊕ H2 , and relative to this decomposition every T ∈ B(H) has the form T = (Ti j ) where Ti j ∈ B(H j , Hi ). In particular, identifying A with π(A), we have that a11 a12 : ai j ∈ Ai j , A= a21 a22 where Ai j = ei Ae j , with Aii ⊆ B(Hi ) unital C ∗ -subalgebras and A21 = A∗12 . The operator space A12 ⊆ B(H2 , H1 ) will be referred to as a corner of A. Note that A11 · A12 · A22 ⊆ A12 , so that A12 is an A11 –A12 -bimodule. Returning to I (S X ), relative to e1 and e2 , we wish to identify each of these four subspaces. Note that X ⊆ e1 I (S X )e2 = I (S X )12 . Theorem 16.1. Let X be an operator space, and S X , I (S X ), e1 , and e2 be as above. Then I (S X )11 , I (S X )22 are injective C ∗ -algebras, and I (S X )12 is the injective envelope of X . Proof. Since the maps γi j : I (S X ) → I (S X )i j deﬁned by γi j (a) = ei ae j are completely contractive projections, it follows that I (S X )i j , i, j = 1, 2, are injective operator spaces and so I (S X )11 and I (S X )22 are injective C ∗ -algebras. To prove the ﬁnal claim it will be enough to prove that I (S X )12 is a minimal injective extension of X . To this end, let ϕ: I (S X )12 → I (S X )12 be completely contractive, and suppose ϕ(x) = x for all x ∈ X . Let λ z ) R= ⊆ I (S X ), : λ, µ ∈ C, z, w ∈ I (S X 12 w∗ µ and let : R → R be deﬁned by λ ϕ(z) λ z = . w∗ µ ϕ(w)∗ µ

228

Chapter 16. Abstract Operator Algebras

Then is completely positive and so extends to a completely positive map, which we still denote by , from I (S X ) to I (S X ). Since is the identity on S X , by rigidity of the injective envelope of S X , is the identity on I (S X ) and hence ϕ must be the identity on I (S X )12 . If X ⊆ Y ⊆ I (S X )12 with Y injective, then there would exist a projection ϕ of I (S X )12 onto Y. By the above, ϕ would be the identity on I (S X )12 and hence Y = I (S X )12 . Thus, I (S X )12 is a minimal injective extension of X , i.e., an injective envelope. Using the uniqueness of the injective envelope, we see that if X ⊆ I with I a minimal injective, then there exists a completely isometric isomorphism between I and I (S X )12 that, restricted to X , is the identity. Thus, we may identify I (S X )12 with I (X ), and under this identiﬁcation, I (X ) is a corner of the C ∗ -algebra I (S X ). We deﬁne I11 (X ) = I (S X )11 and I22 (X ) = I (S X )22 . Thus we have the following picture of the C ∗ -algebra I (S X ):

a z : a ∈ I11 (X ), b ∈ I22 (X ), z, w ∈ I (X ) I (S X ) = w∗ b where I11 (X ) and I22 (X ) are injective C ∗ -algebras and I (X ) is an operator I11 (X )-I22 (X )-bimodule. Moreover, the fact that I (S X ) is a C ∗ -algebra means that for z, w ∈ I (X ), 0 z 0 0 z zw ∗ = , w∗ 0 0 w∗ z w∗ 0 and consequently there are natural products z · w ∗ ∈ I11 (X ), w ∗ · z ∈ I22 (X ). It is interesting to note that setting z, w = zw ∗ deﬁnes an I11 (X )-valued inner product that makes I (X ) a Hilbert C ∗ -module over I11 (X ), but we shall not use this additional structure. This rather complete picture of I (S X ) has some immediate consequences. Corollary 16.2 (Youngson). An operator space is injective if and only if it is completely isometrically isomorphic to a corner of an injective C ∗ -algebra. Proof. If X is an injective operator space, then X = I (X ), which is a corner of the injective C ∗ -algebra I (S X ). Conversely, every corner of an injective C ∗ -algebra is the image of a completely contractive projection and hence is injective.

Chapter 16. Abstract Operator Algebras

229

If X is an operator space, then it is clear that all of the spaces Mm,n (X ) are themselves operator spaces. Technically, the matrix-norm structure on Mm,n (X ) comes from the identiﬁcation M p,q (Mm,n (X )) = M pm,qn (X ). We set Cm (X ) = Mm,1 (X ) and Rn (X ) = M1,n (X ) and call these the column and row operator spaces of X , respectively. In the following results, we characterize the maps from X to itself that are given as multiplication by elements of the corners of I (S X ). All products in the following results are deﬁned by the picture of I (S X ) outlined above. Lemma 16.3. Let X be an operator space and let a ∈ I11 (X ). If a · x = 0 for all x ∈ X , then a = 0. Similarly, if b ∈ I22 (X ) and xb = 0 for all x ∈ X , then b = 0. Proof. Without loss of generality we can assume that a ≤ 1. Let p = a ∗ a so that 0 ≤ p ≤ 1, where 1 denotes the unit of I11 (X ). Deﬁne ϕ: I (X ) → I (X ) by ϕ(z) = (1 − p)z. Since 1 − p ≤ 1, ϕ is a completely contractive map. But for x ∈ X we have ϕ(x) = x, and hence by rigidity ϕ(z) = z for all z ∈ I (X ). Hence, p I (X ) = 0. Now let K denote the two-sided ideal in I (S X ) generated by 0p 00 . Since p 0 p I11 (X ) 0 p 0 I11 (X ) p 0 and I (S X ) , I (S X ) = = 0 0 0 0 0 0 0 0 we see that

k 0 : k ∈ K1 , 0 0

K =

where K 1 is the two-sided ideal in I11 (X ) generated by p. For a1 pa2 ∈ K 1 and z ∈ I (X ) we have that, (a1 pa2 )z = a1 p(a2 z) = 0, since a2 z ∈ I (X ). Thus, K 1 = I11 (X ), and the quotient C ∗ -algebra I (S X )/K is isomorphic to I11 (X )/K 1 I (X ) . I (X )∗ I22 (X ) From this it is clear that the quotient map π: I (S X ) → I (S X )/K is a complete order isomorphism on S X . Hence, there is a complete isometry γ : π (S X ) → I (S X ) such that γ ◦ π is the identity on S X .

230

Chapter 16. Abstract Operator Algebras

By the injectivity of I (S X ) we may extend γ to a completely contractive map, which we still denote by γ , from I (S X )/K to I (S X ). However, since γ ◦ π is the identity on S X , by rigidity it must be the identity on all of I (S X ). But this implies that π is one-to-one and hence K = (0). Therefore, p = 0 and so a = 0. The proof for b ∈ I22 (X ) is similar. Theorem 16.4. Let X be an operator space, let ϕ: X → X be completely bounded, and deﬁne τ : C2 (X ) → C2 (X ) via ϕ(x1 ) x1 τ . = x2 ϕcb x2 Then τ cb = ϕcb if and only if there exists a unique a ∈ I11 (X ), a = ϕcb , such that ϕ(x) = ax for all x ∈ X . Proof. By scaling it is enough to consider the case ϕcb = 1. If such an element a exists, then it must be unique by Lemma 16.3. In this case τ

a 0 x1 x1 = , x2 x2 0 1

and since ( a0 10 ) ≤ 1 in M2 (I11 (X )), it follows easily that τ is completely contractive. So assume that τ is completely contractive, and consider  I11 (X ) I11 (X ) I (X )   B =  I11 (X ) I11 (X ) I (X )  . I (X )∗ I (X )∗ I22 (X ) 

Since  I (X ) I11 (X ) I11 (X ) I (X )  I (X ) I (X ) I (X ) I (X )  11   11 M2 (I (S X )) ∼ = ,  I (X )∗ I (X )∗ I22 (X ) I22 (X ) I (X )∗ I (X )∗ I22 (X ) I22 (X ) 

we see that B is the upper 3 × 3 corner of this injective C ∗ -algebra. Hence B is also an injective C ∗ -algebra, since it is the image of a completely contractive projection applied to an injective C ∗ -algebra .

Chapter 16. Abstract Operator Algebras

231

Let S ⊆ B be the operator system 

C S=0 X∗ It is also useful to think of S as D2 C2 (X ) , C2 (X )∗ C

 0 X C X . X∗ C

where

D2 = C ⊕ C,

and to regard C2 (X ) as a D2 -C-bimodule. Here the left action is given by λ1 0 x1 λ ·x = 1 1 , x2 λ2 · x 2 0 λ2 and the map τ is a D2 -C-bimodule map. If we deﬁne : S → S via     0 ϕ(x1 ) λ1 λ1 0 x 1      0 λ2 x2  =  0 x2  , λ2 y1∗ y2∗ λ3 ϕ(y1 )∗ y2∗ λ3 then since τ is a completely contractive D2 -C-bimodule map, is completely positive. By injectivity we may extend to a completely positive map from B to B, which we still denote by . Since ﬁxes the diagonal C ∗ -subalgebra D3 = C ⊕ C ⊕ C of B, must be a D3 -bimodule map. This implies that there exist maps ϕi, j , i, j = 1, 2, 3, such that ((z i j )) = (ϕi j (z i j )), where the 3 × 3 matrix (z i j ) represents a typical element of B. In particular, ϕ13 : I (X ) → I (X ) is an extension of ϕ: X → X , and ϕ23 : I (X ) → I (X ) is an extension of the identity map on X . Now the map I11 (X ) I (X ) I11 (X ) I (X ) ϕ22 ϕ23 → : ϕ32 ϕ33 I (X )∗ I22 (X ) I (X )∗ I22 (X ) is completely positive and is the identity on ( XC∗ CX ) = S X . Thus, by rigidity it must be the identity on I11 (X ) I (X ) = I (S X ). I (X )∗ I22 (X )

232

Chapter 16. Abstract Operator Algebras

This means that is the identity on the C ∗ -subalgebra   C 0 0  0 I (X ) I (X )  11   0 I (X )∗ I22 (X ) and hence a bimodule map over this C ∗ -algebra. Finally, for x ∈ X we have that      0 0 ϕ(x) 0 0 0 0 1 0   0  = 0 0 0 0 0 x  0 0 0 0 0 0 0 0 0 0 0     0 1 0 0 0 0 = 0 0 0 0 0 x  0 0 0 0 0 0    0 ϕ12 (1) 0 0 0 0   = 0 0 0 0 0 x  0 0 0 0 0 0   0 0 ϕ12 (1)x   = 0 0 0 , 0 0 0 and we see that necessarily ϕ is left multiplication by the element a = ϕ12 (1) ∈ I11 (X ). Since is unital and completely positive, ϕ12 (1) ≤ 1, which completes the proof. Corollary 16.5. Let X be an operator space, let ϕ: X → X be completely bounded, and deﬁne γ : R2 (X ) → R2 (X ) via γ ((x1 , x2 )) = (ϕ(x1 ), ϕcb x2 ). Then γ cb = ϕcb if and only if there exists a unique b ∈ I22 (X ), b = ϕcb , such that ϕ(x) = xb for all x ∈ X . Proof. The proof follows similarly, or by taking adjoints.

We are now in a position to prove the BRS theorem. If A is an algebra, then Mn (A) is also an algebra with product given by (ai j )(bi j ) = ( k aik bk j ). When A is a normed algebra, we call the product contractive provided that ab ≤ ab for every a and b in A. If A is an algebra and an operator space, we call the product completely contractive provided that the product on Mn (A)

Chapter 16. Abstract Operator Algebras

233

is contractive for every n, and we call A an abstract operator algebra. If A also has a unit e and e = 1, then we call A an abstract unital operator algebra. The following theorem shows that every abstract unital operator algebra is in fact, a concrete operator algebra. Theorem 16.6. Let A be an abstract unital operator algebra. Then there exists a unital completely isometric homomorphism π: A → I11 (A). Proof. Let a ∈ A, and let ϕa : A → A be left multiplication by a, i.e., ϕa (x) = ax. Since multiplication is completely contractive, ϕcb ≤ a, but a = ϕ(e) ≤ ϕcb and so ϕcb = a. Consider the map τ : C2 (A) → C2 (A), ϕa (x1 ) a 0 x1 x1 = . = τ x2 x2 0 ae ax2 Since

a 0 0 ae = a

and the multiplication on A is completely contractive, it follows that τ cb = a = ϕcb . Thus, by Theorem 16.4, there is a unique element π (a) ∈ I11 (A), π (a) = a, such that ax = π(a) ◦ x, where we let ◦ denote the action of I11 (A) on I (A). Thus, there is a well-deﬁned map π: A → I11 (A) satisfying π (a) = a and π(a) ◦ x = ax for all x. By Lemma 16.3, it follows that π(a + b) = π(a) + π (b), π(ab) = π(a)π(b), and π (e) = 1, since their differences annihilate A. Hence π is an isometric homomorphism. It remains to show that π is a complete isometry. To see that π is n-contractive, apply the same reasoning to obtain a contractive map β: Mn (A) → I11 (Mn (A)) = Mn (I11 (A)) satisfying β((ai j )) ◦ (xi j ) = (ai j )(xi j ). Then, again by Lemma 16.3, β((ai j )) = (π(ai j )) and hence π is n-contractive, for all n. On the other hand,   e 0 . . .  = (ai j ),  (π(ai j )) ≥ (π(ai j )) ◦ 0 e which shows that π is a complete isometry.

Corollary 16.7 (Blecher–Ruan–Sinclair Theorem). Let A be an abstract unital operator algebra. Then there exists a Hilbert space H and a unital completely isometric homomorphism π: A → B(H).

234

Chapter 16. Abstract Operator Algebras

Proof. Since I11 (A) is a unital C ∗ -algebra, by the Gelfand–Naimark–Segal theorem there exists a Hilbert space H and a unital *-isomorphism of I11 (A) into B(H). Since *-isomorphisms are complete isometries, the result follows.

We are now in a position to generalize the Christensen–Effros–Sinclair representation theorem. Let X be an operator space, and let A and B be unital operator algebras. We call X an operator A-B-bimodule provided that X is an A-B-bimodule and (ai j )(xi j )(bi j ) ≤ (ai j )(xi j )(bi j ) for every m, n, p, q, and every (ai j ) ∈ Mm,n (A), (xi j ) ∈ Mnp (X ), and (bi j ) ∈ M pq (B). When B = C, we call X a left operator A-module, and when A = C, we call X a right operator B-module. It is important to recall that, generally, there is a difference between X being an A-B-bimodule and X being both a left A-module and a right B-module. The difference is that a bimodule action is assumed to be associative, i.e., (a ◦ x) ◦ b = a ◦ (x ◦ b), whereas if X is only a left A-module and a right B-module, then no associativity is assumed. Thus, to be an A-B-bimodule is, generally, a stronger condition. However, as we shall see, in this situation it is all the same. Theorem 16.8. Let X be an operator space, with A and B unital operator algebras. If X is a left operator A-module (respectively, right operator Bmodule), then there is a unique completely contractive homomorphism π : A → I11 (X )(ρ: B → I22 (X )) such that π(a) · x = a ◦ x(x · ρ(b) = x ◦ b), where ◦ denotes the module action and · denotes the product in I (S X ). Proof. As in the proof of Theorem 16.6, for a ∈ A the map a ◦ x1 a 0 x1 x τ = = ◦ 1 x2 x2 0 ae ax2 is completely bounded with τ cb = a and so is given as multiplication by a unique element of I11 (X ). This deﬁnes π : A → I11 (X ). To deﬁne ρ one uses similar arguments on γ ((x1 , x2 )) = (x1 ◦ b, bx2 ) and invokes Corollary 16.5.

Chapter 16. Abstract Operator Algebras

235

One consequence of the above theorem is that the associativity condition for bimodule actions is automatic. Corollary 16.9. Let X be an operator space, and let A and B be unital operator algebras. If X is a left operator A-module and a right operator B-module, then X is an operator A-B-bimodule. Proof. Let π and ρ be as above. Then for any a ∈ A, b ∈ B, and x ∈ X , we have a ◦ (x ◦ b) = π(a)(xρ(b)) = (π(a)x)ρ(b) = (a ◦ x) ◦ b.

The following result generalizes the Christensen–Effros–Sinclair representation theorem (Theorem 15.14). Corollary 16.10. Let A and B be unital operator algebras, and let X be an operator A-B-bimodule. Then there exist Hilbert spaces H and K, unital completely contractive homomorphisms π: A → B(H), ρ: B → B(K), and a complete isometry ϕ: X → B(K, H) such that π(a)ϕ(x)ρ(b) = ϕ(axb). Proof. Represent the C ∗ -algebra I (S X ) as a C ∗ -subalgebra of a Hilbert space L, set H = e1 L, K = e2 L, and apply Theorem 16.8. A representation of an operator A-B-bimodule of the form given by Corollary 16.10 is called a Christensen–Effros–Sinclair representation. Theorem 16.8 is much stronger than Corollary 16.10; it shows that every A-B-bimodule action of operator algebras on an operator space X factors through the actions of I11 (X ) and I22 (X ) on X . To be more precise, let M (X ) = {a ∈ I11 (X ): a X ⊆ X } and Mr (X ) = {b ∈ I22 (X ): X b ⊆ X }; then these sets are easily seen to be unital subalgebras of the C ∗ -algebras I11 (X ) and I22 (X ), respectively. We call these operator algebras, respectively, the left and right multiplier algebras of X . We have that X is an operator M (X )-Mr (X )bimodule. By Theorem 16.8, if X is an operator A-B-bimodule, then there exist completely contractive homomorphisms π: A → M (X ) and ρ: B → Mr (X )

236

Chapter 16. Abstract Operator Algebras

such that a · x · b = π(a)xρ(b). Thus, in this sense, M (X ) and Mr (X ) are the “universal” operator algebras that can act on X . It is also possible to give an “extrinsic” characterization of multipliers. Following Blecher [22], we call a linear map ϕ: X → X a left multiplier provided that there exists a Hilbert space H an operator T in B(H) and a completely isometric map ψ: X → B(H) such that ψ(ϕ(x)) = T ψ(x) for every x in X. When this is the case, we say that (ψ, T ) is an implementing pair for ϕ. Thus, a map is a left multiplier exactly when, after a suitable representation of the operator space X , the map is given as multiplication by an operator. If ϕ is a left multiplier, then we deﬁne its multiplier norm ϕm to be the inﬁmum of T over all such representations. The next result shows that this inﬁmum is always attained. Proposition 16.11. Let X be an operator space, and let ϕ be a left multiplier of X . Then there exists a Hilbert space H, a complete isometry ψ: X → B(H), and an operator T in B(H) such that (ψ, T ) implements ϕ and ϕm = T . Proof. Let (ψn , Tn ) be a sequence of implementing pairs on Hilbert spaces Hn for ϕ with ϕm = limn Tn . Let F be the vector space of bounded sequences of vectors h = (h 1 , h 2 , . . . ) with h n in Hn . Fix a Banach generalized limit glim, and deﬁne a sesquilinear form on F by setting h, k = glimn h n , kn for h = (h 1 , h 2 , . . . ) and k = (k1 , k2 , . . . ) in F. Let H denote the Hilbert space obtained by forming the quotient of F by the null space of this sesquilinear form and completing. If we deﬁne ψ: X → B(H) and deﬁne T in B(H) by setting T h, k = glimn Tn h n , kn and ψ(x)h, k = glimn ψn (x)h n , kn , then it is easily checked that (ψ, T ) is an implementing pair for ϕ with ϕm = T . These “extrinsically” deﬁned maps have the following “intrinsic” characterizations. Theorem 16.12 (Blecher–Effros–Zarikian Theorem). Let X be an operator space, and let ϕ: X → X be a linear map. Then ϕ is a left multiplier with ϕm ≤ 1 if and only if the map τ : C2 (X ) → C2 (X ) deﬁned by ϕ(x1 ) x1 τ = x2 x2 is a complete contraction.

Notes

237

Proof. If τ is a complete contraction, then by Theorem 16.4 there exists an element a ∈ I11 (X ) such that ϕ(x) = ax with a = 1. Representing I (S X ) as operators on a Hilbert space, letting ψ be the induced mapping of X , and letting T be the image of a yields the desired implementing pair. Conversely, if ϕm ≤ 1, then by using an implementing pair (ψ, T ) with ϕm = T it is easy to see that τ is completely contractive. Combining Theorem 16.4 with the above result completes the description of multipliers. Corollary 16.13. Let X be an operator space, and let ϕ: X → X be a linear map. Then ϕ is a left multiplier if and only if there exists a unique a ∈ I11 (X ) such that ϕ(x) = ax. Moreover, in this case ϕm = a. Thus, we see that “extrinsically” deﬁned left multipliers are exactly the maps given as left multiplications by the elements of M (X ). There is a parallel theory for right multipliers.

Notes The original proofs of the BRS theorem and the representation theorem for operator bimodules (i.e., the generalization of the Christensen–Effros–Sinclair representation theorem) made use of a representation theory for completely bounded multilinear maps, which we shall study in the next chapter. In [102] it was shown that the left multiplier algebra of an ideal J of a C ∗ -algebra A can be represented as the subalgebra {x ∈ I (A): x J ⊂ J }. Blecher [23] developed the study of multipliers of operator spaces. Many of these ideas were developed, independently, by Werner [250]. The techniques of [102] were then extended to multipliers of operator spaces in [30], where Theorem 16.8 appeared. But this earlier proof of Theorem 16.8 still used the representation theory for completely bounded multilinear maps in an essential way. Blecher, Effros, and Zarikian [25], in their proof of Theorem 16.12, introduced the use of the map τ . Theorem 16.4 is implied by Theorem 16.12 together with Theorem 16.8, which both appeared earlier. The proof of Theorem 16.4 given here is new and avoids reliance on the representation theory for completely bounded multilinear maps. The key ingredients are the τ -trick of [25] and the methods from [102]. The important role played by I (S X ) for many results about operator spaces was recognized earlier in [204], [113], and [249].

238

Chapter 16. Abstract Operator Algebras Exercises

16.1 Let Cn denote n-dimensional column Hilbert space. Prove that I (SCn ) ∼ = Mn+1 in such a way that I11 (Cn ) = M (Cn ) ∼ = Mn and I22 (Cn ) = Mr (Cn ) ∼ = C. 16.2 Prove, by showing that the BRS axioms are met, that the matrix-normed algebra (Pn , ·u,k ), introduced in Chapter 5, is a unital operator algebra. 16.3 Let A be a unital operator algebra, and let J be a nontrivial two-sided ideal in A. Prove that the algebra A/J equipped with the quotient operator space structure is an operator algebra.

Chapter 17 Completely Bounded Multilinear Maps and the Haagerup Tensor Norm

In many parts of analysis an important role is played by multilinear maps. Recall that if E, F, and Z are vector spaces, then a map γ : E × F → Z is bilinear provided that it is linear in each variable, i.e., γ (e1 + e2 , f 1 ) = γ (e1 , f 1 ) + γ (e2 , f 1 ), γ (e1 , f 1 + f 2 ) = γ (e1 , f 1 ) + γ (e1 , f 2 ), and γ (λe1 , f 1 ) = γ (e1 , λ f 1 ) = λγ (e1 , f 1 ) for any e1 and e2 in E, for any f 1 and f 2 in F, and for any λ in C. If one forms the algebraic tensor product E ⊗ F of E and F, then there is a one-to-one correspondence between linear maps : E ⊗ F → Z and bilinear maps γ : E × F → Z given by setting (e ⊗ f ) = γ (e, f ). Consequently, if one endows E ⊗ F with a matrix norm, then the completely bounded linear maps from E ⊗ F to another matrix-normed space Z correspond to a family of bilinear maps from E × F to Z that one would like to regard as the “completely bounded” bilinear maps. In this fashion, one often arrives at an important family of bilinear maps to study. Conversely, if one starts with a family of bilinear maps into Z that one chooses to regard as the “completely bounded” bilinear maps, then often there is a corresponding matrix-normed structure on E ⊗ F such that these maps are precisely the completely bounded linear maps from E ⊗ F to Z , and in this fashion one often arrives at an important tensor norm. There are several important cases in the literature where operator space norms on tensor products of operator spaces and the corresponding families of bilinear maps are examined in detail. Among these are the spatial tensor norm, also called the operator space injective tensor norm, the operator space projective tensor norm, and the Haagerup tensor norm. We touched on some of these tensor norms in Chapter 12. For the remainder of this book only the Haagerup tensor norm and its corresponding family of “completely bounded” bilinear maps play a crucial role. So we will develop the properties of this norm and the corresponding maps in

239

240

Chapter 17. Completely Bounded Multilinear Maps

some depth. We refer the interested reader to [88] and [193] for a more in-depth study of the tensor theory of operator spaces. The original proofs of the Blecher–Ruan–Sinclair characterization of operator algebras, and the various characterizations of operator bimodules, such as the Christensen–Effros–Sinclair theorem, used the Haagerup tensor theory. In particular, these proofs used a Stinespring-like representation theorem for the corresponding family of bilinear maps, which we shall prove later in this chapter. The Haagerup tensor theory is also closely connected with the theory of free products of C ∗ -algebras. Consequently, we begin this chapter with an analysis of the Haagerup tensor norm followed by some discussion of free products. We then present alternate proofs of several of the results of Chapter 16. In particular, we present proofs of the BRS characterization theorem of operator algebras (Corollary 16.7) and of the characterization theorem for operator bimodules (Corollary 16.10) that use this Stinespring-like representation theorem for the corresponding family of completely bounded bilinear maps. The Haagerup tensor norm also will play a central role in Chapter 19. In a certain sense the Haagerup tensor norm is the “universal norm for products.” To make this clear we introduce a “faux” product operation, denoted . Given an element e of E and an element f of F, we do not necessarily have a way to take their product, but the element e ⊗ f obeys the same bilinear relations as a product. Extending this idea, given (ei j ) ∈ Mm,k (E) and ( f i, j ) ∈ Mk,n (F), we deﬁne an element (ei j ) ( f i j ) of Mm,n (E ⊗ F) by setting k (ei j ) ( f i j ) = ei ⊗ f j . =1

For (xi j ) ∈ Mm,n (E ⊗ F) we deﬁne the Haagerup norm by setting (xi j ) h = inf{ (ei j ) ( f i j ) : (xi j ) = (ei j ) ( f i j )}, where the inﬁmum is taken over all possible ways to represent (xi j ) = (ei j ) ( f i j ) with (ei j ) ∈ Mm,k (E), ( f i j ) ∈ Mk,n (F), and k arbitrary. We let E ⊗h F denote E ⊗ F equipped with this set of norms. It is important to note that the above inﬁmum is nonempty, that is, given (xi j ), there is always a k, a (ei j ) ∈ Mmk (E), and a ( f i j ) ∈ Mkn (F) such that (xi j ) = (ei j ) ( f i j ). To see this, note that if x = k=1 e ⊗ f in E ⊗ F then   f1  ..  x = (e1 , . . . , ek )  .  , fk

Chapter 17. Completely Bounded Multilinear Maps while

 f1  e1 e2 0 0 e1 ⊗ f 1 e 2 ⊗ f 2 0 =   f3 0 0 e3 3 4 e3 ⊗ f 3 e4 ⊗ f 4 0

241

 0 f2  . 0 f4

Combining these two tricks allows one to express an arbitrary (xi j ) as (ei j ) ( f i j ). We wish to show that the Haagerup norm is indeed a norm on E ⊗ F and that E ⊗ F is an L ∞ -matrix-normed space, i.e., an abstract operator space, when endowed with the family of Haagerup norms. To do this a few lemmas are useful. Lemma 17.1. Let (ei j ) ∈ Mm,k (E), ( f i j ) ∈ Mk,n (F), and let A, B, and C be scalar matrices with A = (ai j ) ∈ Mm 1 ,m , B ∈ Mk,k , and C ∈ Mn,n 1 . Then, (i) A · [(ei j ) ( f i j )] = [A(ei j )] ( f i j ), (ii) (ei j )B ( f i j ) = (ei j ) B( f i j ), (iii) [(ei j ) ( f i j )] · C = (ei j ) [( f i j )C].

Proof. Straightforward.

We shall call a family of norms on Mm,n (V ) satisfying all the axioms for an L ∞ -matrix norm except that x = 0 implies x = 0 for x ∈ V an L ∞ -matrix seminorm. If we let W = {x: x = 0}, then these induce an L ∞ -matrix norm on V /W and it becomes an abstract operator space (Exercise 17.1). Proposition 17.2. Let E and F be operator spaces. Then the Haagerup norm is an L ∞ -matrix seminorm on E ⊗ F. Proof. By Lemma 17.1 we have for (xi j ) ∈ Mm,n (E ⊗ F), A ∈ Mm 1 ,m , and B ∈ Mn,n 1 that A(xi j )B h ≤ A (xi j ) h B . Next we show the triangle inequality. Given (xi j ) and (yi j ) in Mm,n (E ⊗ F), and > 0, choose (ei j ) ∈ Mm,k1 (E), ( f i j ) ∈ Mk1 ,n (F), (eˆi j ) ∈ Mm,k2 (E), and ( fˆi j ) ∈ Mk2 ,n (F) such that (xi j ) = (ei j ) ( f i j ), (yi j ) = (eˆi j ) ( fˆi j ), and (ei j ) ( f i j ) ≤ (xi j ) h + , (eˆi j ) ( fˆi j ) ≤ (yi j ) h + .

242

Chapter 17. Completely Bounded Multilinear Maps

Replacing (ei j ), ( f i j ) by r (ei j ), r −1 ( f i j ), we may assume that (ei j ) = ( f i j ) and similarly that (eˆi j ) = ( fˆi j ) . Now consider ( fi j ) ∈ Mk1 +k2 ,n (F). e = ((ei j ), (eˆi j )) ∈ Mm,k1 +k2 (E) and f = ( fˆi j ) We have that (xi j ) + (yi j ) = e f and hence (xi j ) + (yi j ) h ≤ e f ≤ (ei j ) 2 + (eˆi j ) 2 · ( f i j ) 2 + ( fˆi j ) 2 ≤ (xi j ) h + (yi j ) h + 2. To see e 2 ≤ (ei j ) 2 + (eˆi j ) 2 , it is perhaps useful to represent E as a concrete operator space. Since was arbitrary, the triangle inequality follows. Finally, to see the L ∞ condition, note that if (xi j ) = (ei j ) ( f i j ) and (yi j ) = (eˆi j ) ( fˆi j ), then 0 0 0 (ei j ) ( fi j ) (xi j ) = . 0 (yi j ) 0 (eˆi j ) 0 ( fˆi j ) The Haagerup tensor norm has many useful identiﬁcations. We introduce one for future use. Note that if E and F are operator spaces then Cm (E) = Mm,1 (E) is an operator space. Similarly, Rn (E) = M1,n (F) is an operator space. Hence, there is a Haagerup norm on Cm (F) ⊗ Rn (F). Proposition 17.3. Let E and F be operator spaces. Then the map : Cm (E) ⊗h Rn (F) → Mm,n (E ⊗h F) given by

  e1  ...  ⊗ ( f 1 , . . . , f n ) = (ei ⊗ f j ) em 

is an isometry of the Haagerup norm on Cm (E) ⊗ Rn (F) with the Haagerup norm on the m × n matrices over E ⊗ F. Proof. One notes that to compute the Haagerup norm of an element of Cm (E) ⊗ Rn (F), one represents it as a row of, say, k columns, which is really an m × k matrix over E, “times” a column of k rows, which is a k × n matrix over F. Now one checks that preserves this identiﬁcation. Before preceding further in our analysis of the Haagerup norm, it is essential to make a few observations about representations of elements of a

Chapter 17. Completely Bounded Multilinear Maps

243

tensor product. First recall that if E and F are vector spaces and ϕ: E → C is any linear functional, then there is a linear map ϕ ⊗ id: E ⊗ F → F with (e ⊗ f ) = ϕ(e) f . Now suppose that x=

n i=1

ei ⊗ f i =

m

e˜ ⊗ f˜

=1

are two ways to express x and that {e1 , . . . , en } are linearly independent. Choosing functionals ϕi such that ϕi (e j ) = δi j we see that f i = ϕi ⊗ id(x) = m ˜ ˜ ˜ =1 ϕi (e˜ ) f . Hence, span{ f 1 , . . . , f n } ⊆ span{ f 1 , . . . , f m }. Thus, if all the sets are linearly independent in both representations, then span{ f 1 , . . . , f n } = span{ f˜1 , . . . , f˜n }, span{e1 , . . . , en } = span{e˜1 , . . . , e˜m }, and n = m. n Also note that one can always write x = i=1 ei ⊗ f i with {e1 , . . . , en } and { f 1 , . . . , f n } linearly independent. For if, say, en = λ1 e1 + · · · + λn−1 en−1 , n−1 ei ⊗ ( f i + λi f n ) and we proceed inductively. The unique inthen x = i=1 n ei ⊗ f i with both {e1 , . . . , en } and { f 1 , . . . , f n } teger n such that x = i=1 linearly independent is called the rank of the tensor x, denoted rank(x). Those more familiar with Banach space approaches to tensor theory know that n ei ⊗ f i in E ⊗ F is often identiﬁed with a linear map Tx : E ∗ → F x = i=1 n given by Tx (ϕ) = i=1 ϕ(ei ) f i and that rank(x) = dim range(Tx ). The fact that the map Tx is independent of the particular representation of x as a sum of elementary tensors is essentially equivalent to our observations above. Theorem 17.4. Let E ⊆ E 1 and F ⊆ F1 be operator spaces. Then the inclusion of E ⊗h F into E 1 ⊗h F1 is a complete isometry. Proof. Let x be in E ⊗ F. Write x 0 for its Haagerup norm as an element of E ⊗h F, and x 1 for its Haagerup norm as an element of E 1 ⊗h F1 . We wish to prove that x 0 = x 1 . By the deﬁnition of the Haagerup norm as an inﬁmum, we have that x 1 ≤ x 0 . n ei ⊗ f i with ei ∈ E 1 and f i ∈ F1 . To prove that x 0 ≤ Now let x = i=1 x 1 , it will be enough to prove that   f1  .  x 0 ≤ (e1 , . . . , en )  ..  . fn Note that the rank of x must be the same whether we regard it as an element m of E ⊗ F or of E 1 ⊗ F1 . So let rank(x) = m ≤ n, and write x = i=1 e˜i ⊗ f˜i with {e˜1 , . . . , e˜m } ⊆ E and { f˜1 , . . . , f˜m } ⊆ F.

244

Chapter 17. Completely Bounded Multilinear Maps

If {e1 , . . . , en } and { f 1 , . . . , f n } are linearly independent, then necessarily m = n, {e1 , . . . , en } ⊆ E, { f 1 , . . . , f n } ⊆ F, and hence   f1  .  x 0 ≤ (e1 , . . . , en )  ..  . fn Thus, we can assume that there are some linear dependences. Say {e1 , . . . , ek } is a maximal linearly independent set of {e1 , . . . , en }. The linear equations expressing {e1 , . . . , en } as linear combinations of {e1 , . . . , ek } can be written as (e1 , . . . , en ) = (e1 , . . . , ek )B for some k × n scalar matrix B. Polar-decompose B = P W with W a k × n partial isometry satisfying W W ∗ = Ik . We have that   f1  ..  x = (e1 , . . . , ek )P W  .  fn   ˆ f1 f1  ..   ..  = (e1 , . . . , ek )P W  .  = (eˆ1 , . . . , eˆk )  .  , fn fˆk where

  ˆ f1 f1  ..   ..  (eˆ1 , . . . , eˆk ) = (e1 , . . . , ek )P and  .  = W  .  . fˆk fn

Note that (eˆ1 , . . . , eˆk ) = (e1 , . . . , en ) , and that     fˆ1 f 1 .  .   ..   .  ≤  .  . fˆ f n k Thus, we have that

    fˆ1 f1 .  .   .  (eˆ1 , . . . , eˆk )  .  ≤ (e1 , . . . , en )  ..  . fˆ fn k

Now repeat this process if necessary on { fˆ1 , . . . , fˆk } until one achieves a situation where both sets {eˆ1 , . . . , eˆk } and { fˆ1 , . . . , fˆk } are linearly independent

Chapter 17. Completely Bounded Multilinear Maps

245

and hence lie in E and F, respectively. Since these elements lie in E and F, one has     fˆ1 f1 .  .   .  x 0 ≤ (eˆ1 , . . . , eˆk )  .  ≤ (e1 , . . . , en )  ..  fˆ fn k

and hence x 0 ≤ x 1 . This proves that the inclusion of E ⊗h F into E 1 ⊗h F1 is isometric. To prove that it is a complete isometry, one notes that Cm (E) ⊆ Cm (E 1 ), Rn (F) ⊆ Rn (F1 ) and applies the result just proven to conclude that the inclusion of Cm (E) ⊗h Rn (F) into Cm (E 1 ) ⊗h Rn (F1 ) is isometric. Now, applying Proposition 17.3, we conclude that Mm,n (E ⊗h F) = Cm (E) ⊗h Rn (F) ⊆ Cm (E 1 ) ⊗h Rn (F1 ) = Mm,n (E 1 ⊗h F1 ) isometrically. Hence the inclusion of E ⊗h F into E 1 ⊗h F1 is a complete isometry.

Corollary 17.5. Let E and F be operator spaces, and let x ∈ E ⊗ F be a rank k ei ⊗ f i such that k tensor. Then there exists a representation x = i=1   f1  .  x h = (e1 , . . . , ek )  ..  . fk In particular, x h = 0, and so the Haagerup norm is indeed a norm. Proof. By the proof of Theorem 17.4, the expression in the Haagerup norm is reduced by making the terms used in the expression of x as a sum of elementary tensors linearly independent. Thus, if rank(x) = k, then     f1   k    ..  ei ⊗ f i . x h = inf (e1 , . . . , ek )  .  : x =     i=1 fk But if x=

k i=1

ei ⊗ f i =

k

e˜i ⊗ f˜i ,

i=1

then span{e1 , . . . , ek } = span{e˜1 , . . . , e˜k } = E 1 and span{ f 1 , . . . , f k } = span{ f˜1 , . . . , f˜k } = F1 .

246

Chapter 17. Completely Bounded Multilinear Maps

Thus, one can choose sequences of bases for E 1 , {e1n , . . . , ekn }, and for k ein ⊗ f in and F1 , { f 1n , . . . , f kn }, such that x = i=1  n  f 1 n 1 1  ..  1/2 1/2 n e , . . . , e ≤ x + ,  .  ≤ x h + . 1 k h n n fn k

Since E 1 and F1 are ﬁnite-dimensional, these bounded sequences will have a common convergent subsequence. Letting e1 , . . . , ek , f 1 , . . . , f k denote their limits, we have that   f1 k  .  ei ⊗ f i with (e1 , . . . , ek )  ..  = x h . x= i=1 fk Since rank(x) = k, the sets {e1 , . . . , ek } and { f 1 , . . . , f k } are bases. Thus, ei = 0, f j = 0 for all i and j, and hence   f1  .  x h = (e1 , . . . , ek )  ..  = 0. fk Combining Proposition 17.2 with Corollary 17.5 establishes that the Haagerup tensor product of two operator spaces is again an operator space. We now wish to give a more concrete realization of the Haagerup tensor product. Since operator spaces E and F can be represented as subspaces of C ∗ -algebras A and B, and since E ⊗h F ⊆ A ⊗h B completely isometrically, it will be sufﬁcient to understand A ⊗h B more clearly when A and B are C ∗ -algebras. To this end we introduce the free product of algebras. Given two algebras A and B, their free product [10] is another algebra, denoted A ∗ B, that contains A and B as subalgebras and satisﬁes the following universal property: If C is any algebra, and π: A → C and ρ: B → C are algebra homomorphisms, then there exists a unique algebra homomorphism γ : A ∗ B → C with γ (a) = π (a) and γ (b) = ρ(b). The map γ is denoted π ∗ ρ. Alternatively, one can regard A ∗ B as consisting of linear combinations of words in A and B, A ∗ B = span{a1 , b1 , a2 ∗ b2 , b3 ∗ a3 , a4 ∗ b4 ∗ a5 , . . . }, where the ai ’s and bi ’s are arbitrary elements of A and B, respectively, subject to certain relations. Among these relations we would have that the product

Chapter 17. Completely Bounded Multilinear Maps

247

w1 w2 , for a word w1 ending in a different letter than w2 begins with, is simply their concatenation, while, for example, the product of w1 = a1 ∗ b1 with w2 = b2 ∗ a2 is w1 w2 = a1 ∗ (b1 b2 ) ∗ a2 . Also, these words behave multilinearly in each variable, e.g., a1 ∗ b1 ∗ a2 + a1 ∗ b2 ∗ a2 = a1 ∗ (b1 + b2 ) ∗ a2 and a1 ∗ (λb1 ) ∗ a2 = (λa1 ) ∗ b1 ∗ a2 = a1 ∗ b1 ∗ (λa2 ) for any scalar λ. When A and B both contain a common subalgebra C, then one can also form the free product amalgamated over C, denoted A ∗C B, which satisﬁes the following universal property: If D is any algebra, and π : A → D and ρ: B → D are algebra homomorphisms with π(c) = ρ(c) for every c in C, then there exists a unique algebra homomorphism γ : A ∗C B → D with γ (a) = π (a) and γ (b) = ρ(b). One can also give A ∗C B a deﬁnition in terms of words where one allows elements of C to commute with concatenation – for example, (ac) ∗ b = a ∗ (cb). Now assume that A and B are unital C ∗ -algebras (we will consider operator algebras later) that both contain the algebra C = C · 1. We endow A ∗C B with a seminorm by setting x = sup π ∗ ρ(x) , where the supremum is over all Hilbert spaces H and all unital ∗-homomorphisms π : A → B(H) and ρ: B → B(H). Clearly, A ∗C B is a ∗-algebra and x ∗ x = x 2 . If J = {x: x = 0}, then J is a two-sided ∗-ideal, and (A ∗C B)/J is a pre-C ∗ -algebra. We denote its completion by A ∗1 B. It turns out that J = {0}, but we shall not need this harder fact here. See [10] for a proof of it and more on representations of A ∗1 B. This construction allows us to introduce another tensor (semi)norm on operator spaces. If E ⊆ A and F ⊆ B are concrete operator subspaces, then there is a linear map γ : E ⊗ F → A ∗1 B deﬁned by setting γ (e ⊗ f ) = e ∗ f and extending linearly. For (xi j ) ∈ Mm,n (E ⊗ F) we deﬁne (xi j ) 1 = (γ (xi j )) Mm,n (A∗1 B) and shall refer to this as the induced amalgamated free product norm. Without citing [10], it is only apparent that · 1 is a seminorm. But we shall prove that it is equal to the Haagerup norm and, consequently, give an independent proof that it is a norm. More importantly, this will also prove that · 1 is independent of the particular representations of E and F as subspaces of C ∗ -algebras. There is a third way to deﬁne a tensor norm on operator spaces, which we wish to prove equals the Haagerup tensor norm. This deﬁnition was ﬁrst introduced in [170], where it is called the Brown tensor norm, but has since come to be widely known as the factorization norm.

248

Chapter 17. Completely Bounded Multilinear Maps

To deﬁne this tensor norm, ﬁrst suppose that we are given three Hilbert spaces, H1 , H2 , and H3 , and linear maps ϕ: E → B(H2 , H3 ) and ψ: F → B(H1 , H2 ). Then there is a linear map ϕ · ψ: E ⊗ F → B(H1 , H3 ) deﬁned by setting ϕ · ψ(e ⊗ f ) = ϕ(e)ψ( f ) and extending linearly. Thus, we may deﬁne the factorization (semi)norm of (xi j ) in Mm,n (E ⊗ F) by setting (xi j ) f = sup (ϕ · ψ(xi j )) , where the supremum is taken over all H1 , H2 , H3 and all pairs of completely contractive maps ϕ: E → B(H2 , H3 ), ψ: F → B(H1 , H2 ). Since the C ∗ -algebra A ∗1 B can be represented on a Hilbert space and since the inclusions E ⊆ A and F ⊆ B yield completely contractive inclusions ϕ: E → A ∗1 B and ψ: F → A ∗1 B, we see that the map γ = ϕ · ψ used to deﬁne the induced amalgamated free product (semi)norm is just one example of a map ϕ · ψ. Hence, we have that (xi j ) 1 ≤ (xi j ) f for all (xi j ) in Mm,n (E ⊗ f ). Now assume that ϕ and ψ are as in the deﬁnition of the factorization norm and that (xi j ) ∈ Mm,n (E ⊗ F) is represented as (xi j ) = (ei j ) ( f i j ) with (ei j ) ∈ Mm,k (E), ( f i j ) ∈ Mk,m (F). Then we have that k k (ϕ · ψ(xi j )) = ϕ · ψ eik ⊗ f k j ϕ(eik )ψ( f k j ) = =1

=1

= (ϕ(ei j ))(ψ( f i j )). Hence, (ϕ · ψ(xi j )) ≤ (ϕ(ei j )) (ψ( f i j )) ≤ (ei j ) ( f i j ) . Taking the inﬁmum over all factorizations and then the supremum over all ϕ and ψ yields (xi j ) 1 ≤ (xi j ) f ≤ (xi j ) h . We now wish to prove that (xi j ) h ≤ (xi j ) 1 , so that all of these matrix norms are equal. Theorem 17.6 (Christensen–Effros–Sinclair–Pisier). Let A and B be unital C ∗ -algebras, and let E ⊆ A, F ⊆ B be subspaces. Then the map γ : E ⊗h F → A ∗1 B given by γ (e ⊗ f ) = e ∗ f is a complete isometry.

Chapter 17. Completely Bounded Multilinear Maps

249

Proof. By Theorem 17.4, it will be sufﬁcient to prove the case E = A, F = B. To this end we show that A ⊗h B is an operator A-B-bimodule. Deﬁne an A-B-bimodule action on A ⊗h B by setting a · (c ⊗ d) · b = (ac) ⊗ (db). We wish to show that this module action is completely contractive. To see this, note that if (xi j ) ∈ Mm,n (A ⊗ B), (ai j ) ∈ Mm 1 ,m (A), (bi j ) ∈ Mn,n (B), and (xi j ) = (ci j ) (di j ), then (ai j ) · (xi j ) · (bi j ) = [(ai j )(ci j )] [(di j )(bi j )]. Hence, (ai j ) · (xi j ) · (bi j ) h ≤ inf{ (ai j )(ci j ) · (di j )(bi j ) : (xi j ) = (ci j ) (di j )} ≤ (ai j ) (bi j ) · (xi j ) h . Thus, by the Christensen–Effros–Sinclair representation theorem (Theorem 15.14 or Corollary 16.10), there exist Hilbert spaces H and K, unital ∗-homomorphisms π: A → B(H), ρ: B → B(K), and a linear complete isometry ψ: A ⊗h B → B(K, H) such that ψ(a · x · b) = π (a)ψ(x)ρ(b). k ai ⊗ bi we have Let T = ψ(1 ⊗ 1). Then T = 1, and for any x = i=1 that ψ(x) =

k

π(ai )Tρ(bi ).

i=1

Now assume that the spaces H and K are large enough so that there exist unital ∗-homomorphisms π1 : A → B(K),

ρ1 : B → B(H).

If this is not the case, then one can always replace (π, ρ, ψ) by a representation such that it is (Exercise 17.2). Let π˜ = π ⊕ π1 , ρ˜ = ρ1 ⊕ ρ, and let √ 1 − TT∗ T √ , U= −T ∗ I − T ∗T so that U is unitary. We then have that k i=1

π(a ˜ i )U ρ(b ˜ i) =

∗ ψ(x) . ∗ ∗

250

Chapter 17. Completely Bounded Multilinear Maps

∗ , so that Let ρ(b) ˆ = U ρ(b)U ˜ k k π(a ˜ i )U ρ(b ˜ i ) = π˜ (ai )ρ(b ˆ i ) x h = ψ(x) ≤ i=1 i=1 k = π˜ ∗ ρˆ ai ∗ bi ≤ x 1 . i=1

The same calculation applies for a matrix (xi j ), and so we have that γ is a complete isometry. A bilinear map ϕ is called completely bounded if and only if there exists a constant C such that k ϕ(ei , f j ) ≤ C (ei j ) ( f i j ) =1 for every n, k, m and every (ei j ) ∈ Mnk (E), ( f i j ) ∈ Mkm (F). The least such constant C is denoted ϕ cb . Of course, as discussed at the beginning of this chapter, every linear map ψ on E ⊗ F corresponds to a bilinear map ϕ on E × F by simply setting ψ(e ⊗ f ) = ϕ(e, f ). It is easily checked (Exercise 17.3) that ψ is completely bounded on E ⊗h F if and only if ϕ is a completely bounded bilinear map, and that the norms coincide. Thus, completely bounded bilinear maps are precisely the family of bilinear maps that correspond to the Haagerup tensor norm in the sense discussed at the beginning of this chapter. In texts that focus on additional norms on tensor products of operator spaces, the completely bounded bilinear maps corresponding to the Haagerup tensor product are sometimes called the multiplicatively completely bounded maps or the completely bounded maps in the sense of Christensen and Sinclair, in order to avoid confusion. The above theorem leads to a representation theorem for these maps. Corollary 17.7. Let A, B be unital C ∗ -algebras, let E ⊆ A, F ⊆ B be subspaces, and let H be a Hilbert space. Then a bilinear map ϕ: E × F → B(H) is completely bounded if and only if there exists a Hilbert space K, unital ∗homomorphisms π: A → B(K), ρ: B → B(K), and linear maps V, W : H → K such that ϕ(e, f ) = V ∗ π(e)ρ( f )W . Moreover, such a representation can be chosen with ϕ cb = V W . Proof. The linear map ψ: E ⊗h F → B(H) with ψ(e ⊗ f ) = ϕ(e, f ) is completely bounded with ψ cb = ϕ cb . Regarding E ⊗h F ⊆ A ∗1 B completely

Chapter 17. Completely Bounded Multilinear Maps

251

isometrically, we may extend ψ to a completely bounded map on A ∗1 B of the same cb norm. The result now follows from the generalized Stinespring representation of completely bounded maps and the fact that π ∗ ρ is a ∗-homomorphism of A ∗1 B. It is often convenient to think of the above result as a “factorization” theorem for bilinear maps that are completely bounded. The following result makes this more explicit. Corollary 17.8. Let E and F be operator spaces, and let H be a Hilbert space. Then a bilinear map γ : E × F → B(H) is completely contractive if and only if there exists a Hilbert space K and completely contractive maps ψ: F → B(H, K) and ϕ: E → B(K, H) such that γ (e, f ) = ϕ(e)ψ( f ) for every e in E and f in F. Proof. If γ is completely contractive, then by Corollary 17.7 we may write γ (e, f ) = V ∗ π (e)ρ( f )W with V ≤ 1 and W ≤ 1. Setting ϕ(e) = V ∗ π(e) and ψ( f ) = ρ( f )W yields the desired factorization. Conversely, if γ (e, f ) = ϕ(e)ψ( f ) then γ (eik , f k j ) = ϕ(eik )ψ( f k j ) k k = (ϕ(ei j )) · (ψ( f i j )) ≤ (ei j ) ( f i j ) for arbitrary matrices (ei j ) in Mmp (E) and ( f i j ) in M pn (F).

We are now prepared to present another proof of the abstract characterization of operator algebras. First we recall the construction of an inverse limit of Hilbert spaces. Assume that we are given Hilbert spaces {Mk }k≥1 and contraction operators Tk : Mk+1 → Mk . We call a sequence (m 1 , m 2 , . . . ) with m k in Mk coherent provided Tk (m k+1 ) = m k , and bounded provided supk m k is ﬁnite. The set M of bounded, coherent sequences with (m 1 , m 2 , . . . ) = supk m k is a Banach space, denoted M = limk (Mk , Tk ). Since every Tk is a contraction, m k = Tk (m k+1 ) ≤ m k+1 and so supk m k = limk m k . We claim M is a Hilbert space. To see this one checks that the parallelogram identity holds.

252

Chapter 17. Completely Bounded Multilinear Maps

Indeed, for x = (x1 , . . . ) and y = (y1 , . . . ) in M, x + y 2 + x − y 2 = lim( xk + yk 2 + xk − yk 2 ) k

= lim(2 xk 2 + 2 yk 2 ) = 2 x 2 + 2 y 2 . k

The following theorem is a restatement of Corollary 16.7. The proof borrows from ideas in [137]. Theorem 17.9 (Blecher–Ruan–Sinclair Theorem). Let A be a unital algebra with unit e, and assume that A is an (abstract) operator space. If e = 1 and the bilinear map m: A × A → A given by m(a, b) = ab is completely contractive, then there exists a Hilbert space M and a unital completely isometric algebra homomorphism ρ: A → B(M). Proof. Let α0 : A → B(H0 , H1 ) be some linear completely isometric map of the operator space A. Applying Corollary 17.8, we have the completely contractive bilinear map γ (a, b) = α0 (ab) = β1 (a)α1 (b) where α1 : A → B(H0 , H2 ) and β1 : A → B(H2 , H0 ) are completely contractive. Since (ai j ) = (α0 (ai j e)) = (β1 (ai j )α1 (e)) ≤ (β1 (ai j )) we see that β1 , and similarly α1 , are complete isometries. Inductively, we obtain complete isometries αk : A → B(H0 , Hk+1 ), βk : A → B(Hk+1 , Hk ) such that αk−1 (ab) = βk (a)αk (b). Now let Mk ⊆ Hk , k ≥ 1, denote the closed linear span of vectors of the form αk−1 (a)h for a in A and h in H0 . Since βk (e)αk (a)h = αk−1 (a)h, we have that βk (e): Mk+1 → Mk . Thus, we may form the Hilbert space M = limk (Mk , βk (e)). If (m 1 , m 2 , . . . ) is a bounded, coherent sequence, then we claim that for any a in A, (β1 (a)m 2 , β2 (a)m 3 , . . . ) is another bounded coherent sequence. Clearly, this sequence is bounded. To see that it is coherent note that if m k+2 = i αk+1 (bi )h i , then βk+1 (a)m k+2 = i αk (abi )h i ∈ Mk+1 and so βk (e)αk (abi )h i = αk−1 (abi )h i βk (e)[βk+1 m k+2 ] = i

= βk (a)

i

αk (bi )h i = βk (a)βk+1 (e)m k+2 = βk (a)m k+1 .

i

Note that we have also shown that βk (e)βk+1 (a) = βk (a)βk+1 (e) on Mk+2 .

Chapter 17. Completely Bounded Multilinear Maps

253

Thus, there is a well-deﬁned map ρ: A → B(M) given by ρ(b)(m 1 , m 2 , . . . ) = (β1 (b)m 2 , β2 (b)m 3 , . . . ). Note that ρ(e) = IM since βk (e)m k+1 = m k . To verify ρ(ab) = ρ(a)ρ(b) it is enough to compare the kth entry of any vector. If m = (m 1 , m 2 , . . . ) is in M and m k+2 = i αk+1 (ci )h i , then the kth entry of ρ(a)ρ(b)m is βk (a)βk+1 (b)m k+2 = βk (a) βk+1 (b)αk+1 (ci )h i = βk (a) αk (bci )h i =

i

αk−1 (abci )h i =

i

i

βk (ab)αk (ci )h i = βk (ab)m k+1 ,

i

which is the kth entry of ρ(ab)m. To complete the proof, we must show that ρ is a complete isometry. Note that to compute (ρ(ai j )) we must apply this matrix to a column of vectors from M, but the norm of each such vector is obtained by taking its norm in Mk and taking a limit on k. Moreover these norms are increasing with k. Thus, if v j = (m 1, j , m 2, j , . . . ), j = 1, . . . , n, are in M and (ai j ) ∈ Mm,n (A), then     m k+1,1 v 1  .  (ρ(ai j ))  ..  . (β = lim (a ))   k i j . . k v m n

M(m)

k+1,n

Mk(m)

  m k+1,1  .  ≤ lim (βk (ai j ))  ..  k m k+1,n   v1 .   = (ai j ) .. . v n

Mk(m)

Hence, (ρ(ai j )) ≤ (ai j ) , and ρ is completely contractive. Finally, since α0 is a complete isometry, given any > 0 and (ai j ) ∈ Mm,n (A), there exists h 1 , . . . , h n in H0 , with h 1 2 + · · · + h n 2 = 1, such that   h1  ..  (ai j ) − ≤ (α0 (αi j ))  .  . hn Forming the coherent sequences vi = (α0 (e)h i , α1 (e)h i , . . . ), we have that

254

Chapter 17. Completely Bounded Multilinear Maps

vi ≤ h i and

    αk (e)h 1 v 1  ..   ..  = lim (ρ(a (β (ρ(ai j )) ≥ )) (a ))   i j k i j . . k vn αk (e)h n     α1 (e)h 1 h1  ..   ..  ≥ (β1 (ai j ))  .  = (α0 (ai j ))  .  . hn α1 (e)h n

Since > 0 was arbitrary, (ρ(ai )) ≥ (ai j ) and we have that ρ is a complete isometry. Note that by Exercise 17.3 the requirement that the multiplication map m: A × A → A be completely contractive is just the requirement that ( k=1 ai bj ) ≤ (ai j ) (bi j ) for all (ai j ) ∈ Mm,k (A), all (bi j ) ∈ Mk,n (A), and all m, k, n. However, by ﬁlling matrices with 0’s to make them square it is easily seen that it is sufﬁcient to consider the cases m = k = n. In these cases, {Mn (A)}n≥1 are all algebras, and the above condition is just the requirement that the product in these algebras satisfy (ai j )(bi j ) ≤ (ai j ) (bi j ) . Thus, Theorem 17.9 is really just a restatement of Corollary 16.7, as claimed earlier. The Haagerup tensor theory has a multilinear generalization. We discuss the trilinear case in enough detail that the multilinear cases should be transparent. Given three operator spaces E, F, G, introduce a “faux” product by setting (ei j ) ( f i j ) (gi j ) = eik ⊗ f k ⊗ gj k,

whenever (ei j ) ∈ Mm,n 1 (E), ( f i j ) ∈ Mn 1 ,n 2 (F), and (gi j ) ∈ Mn 2 ,n (G). Using this operation, we deﬁne for each m, n a norm on Mm,n (E ⊗ F ⊗ G) by setting (xi j ) h = inf{ (ei j ) ( f i j ) (gi j ) : (xi j ) = (ei j ) ( f i j ) (gi j )}, where the inﬁmum is over all n 1 , n 2 and all such representations of (xi j ). We let E ⊗h F ⊗h G denote the resulting matrix-normed space. On the other hand E ⊗h F and F ⊗h G are operator spaces, and so we can form (E ⊗h F) ⊗h G and E ⊗h (F ⊗h G). Proposition 17.10. Let E, F, and G be operator spaces. Then the linear isomorphisms ϕ: (E ⊗h F) ⊗h G → E ⊗h F ⊗h G and ψ: E ⊗h (F ⊗h G) → E ⊗h F ⊗h G given by ϕ((e ⊗ f ) ⊗ g) = e ⊗ f ⊗ g = ψ(e ⊗ ( f ⊗ g)) are complete isometries.

Chapter 17. Completely Bounded Multilinear Maps

255

Proof. Exercise 17.3.

The above proposition is generally summarized by the statement that the Haagerup tensor product is associative, and given operator spaces E 1 , . . . , E N we write E 1 ⊗h · · · ⊗h E N with no further explanation. In a similar fashion the amalgamated free product is associative. That is, if A1 , A2 , and A3 are unital C ∗ -algebras, then one can deﬁne A1 ∗1 A2 ∗1 A3 and (A1 ∗1 A2 ) ∗1 A3 = A1 ∗1 A2 ∗1 A3 = A1 ∗1 (A2 ∗1 A3 ) ∗-isomorphically, via the natural identiﬁcations. This C ∗ -algebra is the universal C ∗ -algebra for triples of unital ∗-homomorphisms πi : Ai → B(H). Consequently, if A1 , . . . , A N are unital C ∗ -algebras, we write A1 ∗1 · · · ∗1 A N , unambiguously. The following results summarize the multilinear theory. We leave the details of their proofs to the interested reader. Theorem 17.11 (Christensen–Effros–Sinclair–Pisier). Let E 1 , . . . , E N be operator spaces and A1 , . . . , A N unital C ∗ -algebras, and assume that E 1 ⊆ A1 , . . . , E N ⊆ A N . Then the linear map ϕ: E 1 ⊗h · · · ⊗h E N → A1 ∗1 · · · ∗1 A N given by ϕ(e1 ⊗ · · · ⊗ e N ) = e1 ∗ · · · ∗ e N is a complete isometry. Theorem 17.12. Let E 1 , . . . , E N be operator spaces and A1 , . . . , A N unital C ∗ -algebras, and assume that E 1 ⊆ A1 , . . . , E N ⊆ A N . If ϕ: E 1 ⊗h . . . ⊗h E N → B(H) is completely bounded, then there exists a Hilbert space K, unital ∗-homomorphisms πi : Ai → B(K), and operators S: K → H, T : H → K such that ϕ(e1 ⊗ · · · ⊗ e N ) = Sπ1 (e1 ) · · · π N (e N )T with ϕ cb = S T . Just as in the bilinear theory, one deﬁnes a multilinear map γ : E 1 × · · · × E N → B(H) to be completely bounded if and only if it extends to give a completely bounded linear map ϕ: E 1 ⊗h · · · ⊗h E N → B(H) with ϕ(e1 ⊗ · · · ⊗ e N ) = γ (e1 , . . . , e N ), and one sets γ cb = ϕ cb . As in the bilinear case, the completely bounded norm of a multilinear map can be characterized by using “faux” products of matrices. The above theorem is generally referred to as the representation or factorization theorem for completely bounded multilinear maps. We close this chapter with an application of the trilinear theory by proving a result that is equivalent to the generalized Christensen–Effros–Sinclair representation theorem (Corollary 16.10).

256

Chapter 17. Completely Bounded Multilinear Maps

Theorem 17.13. Let A and B be (abstract) unital operator algebras, and let X be an operator space that is also an A-B-bimodule with e A · x · e B = x for all x in X . If the bimodule action (a, x, b) → axb is a completely contractive trilinear map, then there exist Hilbert spaces H and K, unital completely isometric homomorphisms π: A → B(H), ρ: B → B(K), and a complete isometry α: X → B(H, K) such that π(a)α(x)ρ(b) = α(a · x · b). Proof. Consider the set a x C= : a ∈ A, b ∈ B, x ∈ X . 0 b If we deﬁne

and

a1 + a2 x 1 + x 2 a2 x 2 a1 x 1 + = 0 b1 0 b2 0 b1 + b2 a1 a2 a1 x2 + x1 b2 a1 x 1 a2 x 2 · = 0 b1 0 b2 0 b1 b2

then C clearly becomes an algebra. Suppose we can introduce an operator space structure on C such that each of the inclusions of A, B, and X into C are complete isometries and such that C satisﬁes the hypotheses of the BRS theorem. Then it is easily checked that any completely isometric isomorphism γ : C → B(L) will lead to π, ρ, α as above, on setting eA 0 0 0 L, K=γ L, H=γ 0 eB 0 0 a 0 0 0 ρ(b) = PK γ π(a) = PH γ , , and 0 0 0 b H K 0 x α(x) = PH γ . 0 0 K Thus, it remains to show that we can endow C with the desired operator algebra structure. To this end we embed A ⊆ C1 , B ⊆ C3 as unital subalgebras and embed X ⊆ C2 as a subspace, where C1 , C2 , and C3 are unital C ∗ -algebras. This allows us to regard A and B as unital subalgebras of C1 ∗1 C2 ∗1 C3 and identify

Notes

257

A ⊗h X ⊗h B ∼ = span{a ∗ x ∗ b: a ∈ A, x ∈ X, b ∈ B}, completely isometrically. The set a v ˜ C= : a ∈ A, v ∈ A ⊗h X ⊗h B, b ∈ B ⊆ M2 (C1 ∗1 C2 ∗1 C3 ) 0 b is readily seen to form a subalgebra of this C ∗ -algebra. Note that the map ϕ: C˜ → C deﬁned by a ai xi bi a ai ∗ xi ∗ bi = ϕ 0 b 0 b is a homomorphism. We endow C with the quotient operator induced by this map, that is, a x a v a = inf a v : ϕ = 0 b 0 b 0 b 0

space structure x b

.

It is now easily checked that C endowed with this norm satisﬁes the hypotheses of the BRS theorem and that each of the inclusions is a complete isometry. Note that when A and B are C ∗ -algebras, the above result reduces to the Christensen–Effros–Sinclair representation theorem.

Notes A version of the representation theorem for multilinear completely bounded maps (Theorem 17.12) was ﬁrst obtained in the case that E 1 , . . . , E N are C ∗ algebras by Christensen and Sinclair [58]. Their proof used Wittstock’s setvalued extension theorem [242]. Later this result was generalized to operator spaces in [170], and the proof was simpliﬁed to the extent that the use of the set-valued extension theorem was replaced by the ordinary extension theorem for linear completely bounded maps. This multilinear representation theorem implied an extension theorem for multilinear completely bounded maps (see Exercise 17.5), which was in turn equivalent to the injectivity of the Haagerup tensor norm (Theorem 17.4). One of the more difﬁcult parts of the proof in [170] was establishing that E 1 ⊗h E 2 had a representation as a concrete operator space, since Ruan’s abstract characterization of operator spaces came later. Indeed, in some ways the proof of that fact in [170] anticipates Ruan’s theorem. The direct proof of the injectivity of the Haagerup tensor norm presented in Theorem 17.4 appeared later in [28].

258

Chapter 17. Completely Bounded Multilinear Maps

The connection between the Haagerup tensor product and free products was ﬁrst made by Christensen, Effros, and Sinclair [57], who proved that the Haagerup tensor product embedded into the unamalgamated free product of the C ∗ -algebras. Pisier [193] was the ﬁrst to realize that the embedding of E 1 ⊗h E 2 into the amalgamated free product is completely isometric (see [162]), which leads to the cleaner realization of the Haagerup tensor product that appears here (Theorem 17.6). However, both of these earlier proofs used either the Christensen–Sinclair representation theorem for completely bounded multilinear maps or its generalization to operator spaces. The original proof of the Christensen–Effros–Sinclair representation theorem for operator bimodules over C ∗ -algebras (Theorem 15.14) also used the Christensen–Sinclair representation theorem for completely bounded bilinear maps. The proofs given here are new in that we have used injective envelopes to directly obtain the Christensen–Effros–Sinclair representation theorem for operator bimodules and the embedding of A1 ⊗h A2 into A1 ∗1 A2 , without appealing to the Christensen–Sinclair representation theorem for completely bounded bilinear maps. Using these results, we are then able to deduce the Christensen–Sinclair representation theorem for bilinear maps and its generalization as a consequence of this embedding. The Haagerup tensor norm ﬁrst appeared in some uncirculated work of Haagerup. Effros and Kishimoto [81] recognized its value and adopted the name. This tensor norm was clearly inspired by Haagerup’s proof [107] of the Wittstock extension theorem for completely bounded linear maps and his results on Schur product maps. Indeed, Wittstock’s extension theorem can be deduced from the injectivity of the Haagerup tensor norm, and this approach recaptures the essence of Haagerup’s proof. We outline how this can be done. One ﬁrst shows that for an operator space E there is an isometric identiﬁcation between CB(E, Mn ) and the dual of Rn ⊗h E ⊗h Cn . Now, let E ⊂ F be operator spaces, so that by the injectivity of the Haagerup tensor norm, Rn ⊗h E ⊗h Cn ⊂ Rn ⊗h F ⊗h Cn , isometrically. Thus, by applying the ordinary Hahn–Banach extension theorem for linear functionals to this latter pair of spaces, one may extend a completely bounded map from E into Mn to a completely bounded map from F into Mn . The details of this approach appear in [28]. See also [162, Theorem 3]. Exercise 17.9 is a special case of the results in [85].

Exercises 17.1 Let V be a vector space, and let · m,n be a family of norms on Mm,n (V ) that are L ∞ -matrix seminorms. Let W = {x ∈ V : |x = 0}. Prove that

Exercises

259

for (xi, j ) ∈ Mm,n (V ), we have (xi, j ) m,n = 0 if and only if (xi, j ) ∈ Mm,n (W ). Conclude that setting (xi, j + W ) m,n = (xi, j ) m,n gives a well-deﬁned L ∞ -matrix norm on V /W. 17.2 Verify the claim in the proof of Theorem 17.6. 17.3 Verify that a bilinear map ϕ: E × F → Z is completely bounded if and only if the induced linear map ψ: E ⊗h F → Z is completely bounded and that ψ cb is the least constant C satisfying k ϕ(ei , f j ) ≤ C (ei j ) ( f i j ) =1

17.4 17.5 17.6

17.7

17.8 17.9

for every n, k, m and every (ei j ) ∈ Mnk (E), ( f i j ) ∈ Mkm (F). Deduce that if E ⊂ E 1 and F ⊂ F1 are operator spaces and ϕ: E × F → B(H) is a completely bounded bilinear map, then ϕ can be extended to a completely bounded bilinear map ψ: E 1 × F1 → B(H) of the same completely bounded norm. Prove the associativity of the Haagerup tensor product as asserted in Proposition 17.10. Extend Exercise 17.3 to the multilinear case, and deduce an extension theorem for completely bounded multilinear maps into B(H). Let A and B be (abstract) unital operator algebras; let X be an operator space that is also an A-B-bimodule with e A · x · e B = x for all x in X . Use Exercise 17.5 to show directly that the bimodule action (a, x, b) → axb is a completely contractive trilinear map if and only if X is an operator A-B-bimodule. Show directly that Cm ⊗h Rn is completely isometrically isomorphic to Mm,n and that this isomorphism carries the tensor product of the standard basis vectors ei ⊗ e j to the standard matrix units E i, j . Prove directly that Rm ⊗h Cn is isometrically isomorphic to the dual of Mm,n . (Effros–Ruan) Let E and F be ﬁnite-dimensional operator spaces, and let E ∗ and F ∗ denote their duals. Prove that (E ⊗h F)∗ is completely isometrically isomorphic to E ∗ ⊗h F ∗ . Deduce Exercise 17.8 from this result.

Chapter 18 Universal Operator Algebras and Factorization

In this chapter we examine some consequences of the BRS characterization of operator algebras. The main theme of this chapter is that by using the BRS theorem, we are able to give intrinsic formulas, similar to the description of MAX(X ), for certain extrinsically deﬁned universal operator algebra norms. Among the results that we shall obtain as consequences of the theory are Nevanlinna’s theorem characterizing the set of analytic functions that map the disk to the disk in terms of positive deﬁnite functions, and Agler’s recent generalizations of Nevanlinna’s theorem. In addition, we obtain formulas for the norm in the full C ∗ -algebra of a group. We begin with the construction of the full operator algebra of a semigroup. By a semigroup we shall mean a nonempty set S, together with a product S × S → S, (s1 , s2 ) → s1 · s2 , and a unit element e satisfying e · s = s · e = s for every s in S. By the semigroup algebra C[S] we mean the vector space of all ﬁnite linear combinations λi si , λi ∈ C, si ∈ S, equipped with the product

λi si

(λi µ j )(si t j ). µjtj =

Note that e is the unit of this algebra. By a representation of S on H we mean any map π: S → B(H) satisfying π(e) = IH and π (s1 · s2 ) = π(s1 )π (s2 ) for all s1 and s2 in S. Clearly, every representation π induces a unital algebra homomorphism of C[S] into B(H), which we still denote by π , on setting π( λi si ) = λi π(si ). We call π bounded if there is a constant C with π (s) ≤ C for all s ∈ S, and contractive when C ≤ 1. If S = Z+ , the nonnegative integers with addition for the product and e = 0 for the unit, then a representation π of Z+ simply corresponds to an operator T = π(1), since π(n) = π(1 + · · · + 1) = π (1)n = T n . The representation

260

Chapter 18. Universal Operator Algebras and Factorization

261

is bounded when T is power-bounded, and contractive when T ≤ 1. The semigroup algebra C[Z+ ] can be identiﬁed with the algebra P of polynomials in one variable. The map γ : C[Z+ ] → P, γ ( λn · n) = λn z n , is an algebra isomorphism between C[Z+ ] and the algebra of polynomials in one variable. Similarly, if (Z+ ) N denotes the Cartesian product of N copies of Z+ , then a contractive representation π: (Z+ ) N → B(H) is determined by choosing an N -tuple of commuting contractions {T1 , . . . , TN } and setting π ((k1 , . . . , k N )) = T1k1 · · · TNk N . The semigroup algebra C[(Z+ ) N ] is isomorphic to P N , the algebra of polynomials in N variables, via the isomorphism γ : C[(Z+ ) N ] → P N deﬁned by γ ((k1 , . . . , k N )) = z 1k1 · · · z kNN . In this chapter we focus primarily on contractive representations. Note that every semigroup possesses at least one contractive representation, given by setting π(s) = IH for all s. We now wish to endow Mm,n (C[S]) with a seminorm by setting (ai j ) = sup(π (ai j )), where the supremum is over all contractive representations π : S → B(H) and all Hilbert spaces. It is easy to see that J = {a ∈ C[S]: a = 0} is a two-sided ideal and that the above formulas deﬁne a matrix norm on C[S]/J that gives C[S]/J the structure of an abstract operator algebra. We denote this algebra by OA(S). We call OA(S) the full semigroup operator algebra and let ·OA(S) denote the norm on OA(S) induced by the seminorm on C[S]. When S = G is a group, then every contractive representation is a unitary representation, since π (g) and π(g)−1 = π (g −1 ) must both be contractions. So in this case the completion of OA(G) is the full group C ∗ -algebra, C ∗ (G), of the discrete group G. By the theorems of von Neumann and Ando, we have that OA(Z+ ) = P is completely isometrically isomorphic to the space of polynomials equipped with the supremum norm over the unit disk, and OA((Z+ )2 ) = P2 is completely isometrically isomorphic to the space of polynomials in two variables equipped with the supremum norm over the bidisk. However, for N ≥ 3 we have that OA((Z+ ) N ) = P N is completely isometrically isomorphic to the algebra of polynomials in N variables equipped with the norm obtained by taking the supremum over all commuting N -tuples of contractions, and this norm is strictly larger than the supremum norm over D N . This is the universal operator algebra for N commuting contractions, (P N , ·u ), which was ﬁrst introduced in Chapter 5. The deﬁnition of the operator algebra structure on OA(S) given above is extrinsic in the sense that it requires a supremum over all representations into

262

Chapter 18. Universal Operator Algebras and Factorization

another object. We now wish to give an intrinsic characterization of this norm, which can be achieved internally to OA(S). Theorem 18.1. Let S be a semigroup and let (ai j ) ∈ Mm,n (C[S]). Then (ai j )OA(S) = inf{C1 · · · C }, where the inﬁmum is taken over all ways to factor (ai j ) = C1 D1 · · · C−1 D−1 C as a product of arbitrarily many matrices over C[S] of arbitrary size, with C1 , . . . , C scalar matrices and D1 , . . . , D−1 diagonal matrices with entries from S. Proof. Let (ai j ) f denote the inﬁmum on the right hand side of the above equation. Note that the set of such factorizations is nonempty. For example, if n λi si is in C[S], then a = i=1    s1 0 . . . 0 1 ..   .. ..  . . . 1 0 a = (λ1 , . . . , λn ) . . . , . .. .. 0   ..  ..  1 0 . . . 0 sn √ and so a f ≤ n(|λ1 |2 + · · · + |λn |2 )1/2 . Given a diagonal matrix   s1 ..  D= . sn as above and a contractive representation π, set   π(s1 )   .. π(D) =  .  π(sn ) and note that π (D) ≤ 1. Given a factorization of (ai j ) as in the statement of the theorem and a contractive representation π, observe that (π(ai j )) = C1 π(D1 ) · · · C−1 π (D−1 )C and hence (π(ai j )) ≤ C1 · · · C . Taking the supremum over all such π and the inﬁmum over all such factorizations yields (ai j )OA(S) ≤ (ai j ) f .

Chapter 18. Universal Operator Algebras and Factorization

263

Assume for the moment that we have proven that (C[S], · f ) satisﬁes the BRS axioms to be an abstract operator algebra seminorm. Then there must exist a Hilbert space H and an algebra homomorphism π : C[S] → B(H) with (ai j ) f = (π(ai j )) for any m, n and (ai j ) in Mm,n (A). Since s f ≤ 1 for s ∈ S, we have that π is a contractive representation of S, and hence (π (ai j )) ≤ (ai j )OA(S) . Thus, (ai j ) f ≤ (ai j )OA(S) , and the result follows. Thus, it remains to check that the axioms of BRS are satisﬁed by the algebra (C[S], · f ). Note that if L and M are scalar matrices of appropriate sizes and (ai j ) = C1 D1 · · · D−1 C is any factorization of the desired type, then L(ai j )M = (LC1 )D1 · · · D−1 (C M) and hence L(ai j )M f ≤ L(ai j ) f M. If (ai j ) = C1 D1 · · · D−1 C and (bi j ) = C1 D1 · · · D −1 C , then (ai j )(bi j ) = C1 D1 · · · · · D−1 (C C1 )D1 · · · D −1 C and hence (ai j )(bi j ) f ≤ C1 · · · C C1 · · · C . Taking the inﬁmum over all such factorizations yields (ai j )(bi j ) f ≤ (ai j ) f (bi j ) f , and so the multiplication is completely contractive. Finally, to see that the triangle inequality and L ∞ condition are met, note that a given factorization can always be made to contain more terms by inserting diagonal matrices of e’s and scalar identity matrices, and the product of the norms remains the same. Thus, given (ai j ) and (bi j ) and > 0, it is enough to assume that we have factorizations (ai j ) = C1 D1 · · · D−1 C and

(bi j ) = C1 D1 · · · D−1 C

with C1 · · · C < (ai j ) f + , C1 · · · C < (bi j ) f + . Replacing Ci by ri Ci with r1 · · · r = 1, we may assume that C1 = · · · = C and C1 = · · · = C . Hence, )(C ⊕ C ) (ai j ) ⊕ (bi j ) = (C1 ⊕ C1 )(D1 ⊕ D1 ) · · · (D−1 ⊕ D−1

and so (ai j ) ⊕ (bi j ) f ≤ C1 ⊕ C1 · · · C ⊕ C ≤ (max{C1 , C1 }) ≤ max{(ai j ) f , (bi j ) f } + , and the L ∞ condition follows. Finally, assuming that (ai j ) and (bi j ) are the same size, to prove the triangle inequality we scale so that C1 2 = C 2 ≤ (ai j ) f + , C1 2 = C 2 ≤

264

Chapter 18. Universal Operator Algebras and Factorization

(bi j ) f + , and C2 = · · · = C−1 = C2 = · · · = C−1 = 1. Then )(D−1 ⊕ D−1 ) (ai j ) + (bi j ) = (C1 , C1 )(D1 ⊕ D1 )(C2 ⊕ C2 ) · · · (C−1 ⊕ C−1 C ( ), and hence C C 2 2 2 2 (ai j ) + (bi j ) f ≤ (C1 , C1 ) C = C1 + C1 C + C

≤ (ai j ) f + (bi j ) f + 2 . Thus, the triangle inequality follows, and we have shown that (C[S], · f ) satisﬁes all the axioms of an operator algebra seminorm. The above theorem implies some factorization results for polynomials and for analytic functions. Corollary 18.2. Let ( pi j (z 1 , . . . , z N )) be a matrix of polynomials in N variables. Then ( pi j (T1 , . . . , TN )) < 1 for all commuting N -tuples of contractions if and only if there exists an integer , scalar matrices Ci with Ci < 1, 1 ≤ i ≤ , and diagonal matrices Di , 1 ≤ i ≤ − 1, of monomials, such that ( pi j (z 1 , . . . , z N )) = C1 D1 · · · D−1 C . Proof. The ﬁrst condition is equivalent to ( pi j )u < 1 in the operator algebra (P N , ·u ). But P N = OA((Z+ ) N ), and the second condition comes from this identiﬁcation and Theorem 18.1. Corollary 18.3. Let ( pi j ) be a matrix of polynomials in N variables, N ≤ 2. Then sup{( pi j (z)): z ∈ D N } < 1 if and only if there exists an integer , scalar matrices Ci with Ci < 1, 1 ≤ i ≤ , and diagonal matrices Di , 1 ≤ i ≤ − 1, of monomials such that ( pi j ) = C1 D1 · · · D−1 C . Proof. Apply Corollary 18.2 and the fact that the universal norm and supremum norm are equal when N ≤ 2. Very little is known about constructing such factorizations. It is natural to wonder if there could exist an integer L such that every polynomial has a factorization as above with ≤ L. The existence of a power-bounded operator that is not polynomially bounded (Theorem 10.9) implies that there cannot exist such an integer L. A proof of this fact is contained in Exercise 18.1. But we do not know a direct argument for the nonexistence of such an L.

Chapter 18. Universal Operator Algebras and Factorization

265

The factorization formula in Corollary 18.2 gives an alternative approach to the questions of Chapter 5, namely, to determining if the operator algebras (P N , ·∞ ) and (P N , ·u ) are boundedly isomorphic. It is also unknown if these two operator algebras are completely boundedly isomorphic. The above results allow us to easily obtain results of Nevanlinna and Agler characterizing polynomials p with pu < 1. To this end we need to introduce the concept of a positive deﬁnite function. Let X be a set, and let H be a Hilbert space. Then a function K : X × X → B(H) is called positive deﬁnite provided that for every ﬁnite subset {x1 , . . . , xn } of X , the operator matrix (K (xi , x j )) is positive deﬁnite. These functions should perhaps be called “completely” positive deﬁnite, but we stay with the classical deﬁnition. The following elementary proposition makes it easy to produce examples of positive deﬁnite functions. Proposition 18.4. Let X be a set, and let H1 , . . . , Hm be Hilbert spaces. (i) If F: X → B(H1 , H2 ) is any function, then K : X × X → B(H1 ) given by K (x, y) = F(x)F(y)∗ is positive deﬁnite. (ii) If K i : X × X → B(Hi ) are positive deﬁnite, then K : X × X → B(H1 ⊕ · · · ⊕ Hn ) given by K (x, y) = K 1 (x, y) ⊕ · · · ⊕ K n (x, y) is positive deﬁnite. (iii) If K i : X × X → B(H1 ), i = 1, . . . , n, are positive deﬁnite, then K 1 + · · · + K n is positive deﬁnite. (iv) If F: X → B(H2 , H1 ) is any function and K : X × X → B(H2 ) is positive deﬁnite, then K 1 : X × X → B(H1 ) deﬁned by K 1 (x, y) = F(x)K (x, y)F(y)∗ is positive deﬁnite. Proof. Exercise 18.2.

A factorization–representation theorem for positive deﬁnite functions implies that for every positive deﬁnite function K : X × X → B(H1 ) there exists a Hilbert space H2 and F: X → B(H1 , H2 ) such that K (x, y) = F(x)F(y)∗ , but we shall not need that here. The scalar-valued version of the following theorem is due to Nevanlinna [153]. Theorem 18.5 (Nevanlinna). Let F = ( f i j ) be an m × n matrix of analytic functions on D. Then sup{F(z): |z| < 1} ≤ 1 if and only if the function

266

Chapter 18. Universal Operator Algebras and Factorization

K : D × D → Mm given by K (z, w) =

I − F(z)F(w)∗ 1 − z w¯

is positive deﬁnite. Proof. If K is positive deﬁnite, then 0 ≤ K (z, z), which implies that F(z) ≤ 1. So assume that F(z) ≤ 1 for all z. It follows from standard techniques in complex analysis that such a function F is a pointwise limit of a sequence of matrices of polynomials {Pn } with sup{Pn (z): |z| < 1} < 1. Assume that we can prove the theorem for such polynomials; then K (z, w) will be the pointwise limit of the sequence of positive deﬁnite functions K n (z, w) = [I − ¯ and hence will be positive deﬁnite. Pn (z)Pn (w)∗ ]/(1 − z w) Thus, it is enough to assume that F = ( pi j ) is a matrix of polynomials with sup{F(z): |z| < 1} < 1. Hence by Corollary 18.3, we may factor F(z) = C1 D1 (z) · · · D−1 (z)C , with C < 1. First assume that F(z) is the constant function C1 . Then K (z, w) = ¯ = (1 − z w) ¯ −1 P, where P = I − C1 C1∗ ≥ 0. But [I − F(z)F(w)∗ ]/(1 − z w) ∞ k k −1 P = k=0 z w¯ P, and each term in this sum is a positive deﬁnite (1 − z w) ¯ function of the form G k (z)G k (w)∗ with G k (z) = z k P 1/2 . Now assume that F(z) = C1 D1 (z). We have that (1 − z w) ¯ −1 (I − F(z)F(w)∗ ) = (1 − z w) ¯ −1 [I − C1 C1∗ + C1 (I − D1 (z)D1 (w)∗ )C1∗ ] = (1 − z w) ¯ −1 (I − C1 C1∗ ) + C1 K (z, w)C1∗ with K (z, w) = (1 − z w) ¯ −1 (I − D1 (z)D1 (w)∗ ). We have already seen that the ﬁrst term in this sum is positive deﬁnite. Note that K (z, w) is a diagonal matrix of functions of the form (1 − z n w¯ n )/ (1 − z w) ¯ = 1 + z w¯ + z 2 w¯ 2 + · · · + z n−1 w¯ n−1 for various integers n. Each term in this sum is positive deﬁnite by Proposition 18.4(i). Further applications of Proposition 18.4 yield that K (z, w) is positive deﬁnite and also that ¯ −1 (I − F(z)F(w)∗ ) is positive C1 K (z, w)C1∗ is positive deﬁnite. Thus, (1 − z w) deﬁnite. The proof now proceeds by induction on . By using an identity matrix if necessary we can always write F(z) = C1 D1 (z) · · · C D (z). Now let F(z) = C1 D1 (z)G(z) where G(z) is of length − 1. Then, (1 − z w) ¯ −1 (I − F(z)F(w)∗ ) = (1 − z w) ¯ −1 (I − C1 D1 (z)D1 (w)∗ C1∗ ) ¯ −1 (I − G(z)G(w)∗ )]D1 (w)∗ C1∗ . + C1 D1 (z)[(1 − z w)

Chapter 18. Universal Operator Algebras and Factorization

267

The ﬁrst term in this sum is positive deﬁnite by the above calculation, and the factor (1 − z w) ¯ −1 (I − G(z)G(w)∗ ) is positive deﬁnite by the inductive hypothesis. Hence, the second term in the sum and, consequently, the sum are positive deﬁnite by Proposition 18.4. Hence, the inductive step follows, and our proof is complete. Theorem 18.6 (Agler). Let F = ( f i j ) be an m × n matrix of analytic functions on D2 . Then sup{F(z 1 , z 2 ): |z 1 | < 1, |z 2 | < 1} ≤ 1 if and only if there exist positive deﬁnite functions Pi : D2 × D2 → Mm such that I − F(z 1 , z 2 )F(w1 , w2 )∗ = (1 − z 1 w¯ 1 )P1 (z 1 , z 2 , w1 , w2 ) + (1 − z 2 w¯ 2 )P2 (z 1 , z 2 , w1 , w2 ). Proof. If such a decomposition exists, then I − F(z 1 , z 2 )F(z 1 , z 2 )∗ ≥ 0 and so F(z 1 , z 2 ) ≤ 1 for all (z 1 , z 2 ). To prove the converse, as in the proof of Theorem 18.5, it is enough to assume that F is a matrix of polynomials and that its supremum is strictly less than 1. Thus, F can be factored as in the statement of Corollary 18.3. The only difference between this factorization and the factorization appearing in the proof of Nevanlinna’s theorem is that the diagonal factors are now monomials in z 1 and z 2 . However, note that 1 − z 1n 1 z 2n 2 w¯ n2 2 w¯ n1 1 = 1 − z 1n 1 w¯ n1 1 + z 1n 1 1 − z 2n 2 w¯ n2 2 w¯ n1 1 = (1 − z 1 w¯ 1 )P1 (z 1 , w1 ) + (1 − z 2 w¯ 2 ) z 1n 1 P2 (z 2 , w2 )w¯ n1 1 , where P1 (z 1 , w1 ) = 1 + (z 1 w¯ 1 ) + · · · + (z 1 w¯ 1 )n 1 −1 and P2 (z 2 , w2 ) = 1 + z 2 w¯ 2 + · · · + (z 2 w¯ 2 )n 2 −1 are positive deﬁnite functions. From this it follows that for each diagonal factor we have that I − D(z 1 , z 2 )D(w1 , w2 )∗ = (1 − z 1 w¯ 1 )Q 1 (z 1 , w1 ) + (1 − z 2 w¯ 2 )Q 2 (z 1 , z 2 , w1 , w2 ) with Q 1 and Q 2 positive deﬁnite. With these observations, the proof can now proceed by induction as in the proof of Theorem 18.5. For functions of N variables, N > 2, we no longer have equality of the supremum norm over D N and the universal norm for N commuting contractions. For this reason Agler introduces the algebra Hu∞ (D N ). A function is in the unit ball of Mm,n (Hu∞ (D N )) if and only if it is the pointwise limit of a sequence of

268

Chapter 18. Universal Operator Algebras and Factorization

matrices of polynomials whose norms in Mm,n ((P N , ·u )) are less than or equal to 1. Thus Au (D N ) ⊆ Hu∞ (D N ) completely isometrically. We also have that Hu∞ (D N ) ⊆ H ∞ (D N ), but this containment is not isometric. However, it is unknown whether or not the containment is strict. In fact it is easily seen (Exercise 18.2) that Hu∞ (D N ) = H ∞ (D N ) if and only if Au (D N ) = A(D N ) if and only if there exists a constant K N such that pu ≤ K N p∞ for all polynomials. This is just another formulation of the open problem raised in Chapter 5: whether or not the N -variable von Neumann inequality holds up to a constant. Theorem 18.6 has been extended to Hu∞ (D N ). However, the “easy” implication is no longer as obvious, and so we only record the result. Theorem 18.7 (Agler). Let F = ( f i j ) be an m × n matrix of functions in Hu∞ (D N ). Then Fu ≤ 1 if and only if there exist positive deﬁnite functions Pi : D N × D N → Mm , 1 ≤ i ≤ N , such that I − F(z)F(w)∗ = (1 − z 1 w¯ 1 )P1 (z, w) + · · · + (1 − z N w¯ N )PN (z, w), where z = (z 1 , . . . , z N ) and w = (w1 , . . . , w N ). The proof that Fu ≤ 1 implies that F has the desired decomposition is similar to the proof of Theorem 18.6 and is left as Exercise 18.3. There are currently few criteria known for when two elements of a semigroup can be separated by a contractive representation. Theorem 18.1 gives us at least one new criterion. Theorem 18.8. Let S be a semigroup and let s1 , s2 be in S. Then π (s1 ) = π (s2 ) for every contractive representation of S if and only if for every > 0 there exists a factorization s1 − s2 = C1 D1 · · · D−1 C such that C1 · · · C < with Ci scalar matrices and Di diagonal matrices of semigroup elements. Proof. We have that π (s1 ) = π(s2 ) for every contractive representation if and only if s1 − s2 OA(S) = 0. By Theorem 18.1, s1 − s2 OA(S) = 0 is equivalent to the latter factorization condition. We now turn our attention to the analogous problem of determining when a normed algebra B with an identity of norm 1 can be represented isometrically as an algebra of operators on a Hilbert space. To this end let Rep(B) denote the collection of unital contractive homomorphisms from B into B(H) for some Hilbert space H. We wish to use Rep(B)

Chapter 18. Universal Operator Algebras and Factorization

269

to deﬁne a “maximal” operator algebra structure on B, which we shall denote MAXA(B), to distinguish it from the maximal operator space structure on the normed space B. For (bi j ) in Mm,n (B) we set (bi j )MAXA(B) = sup{(π(bi j )): π ∈ Rep(B)} when Rep(B) is nonempty, and set it equal to 0 when Rep(B) is empty. Note that if J = {b: π(b) = 0 for all π ∈ Rep(B)}, then J is a two-sided ideal, with J = B when Rep(B) is empty, and that ·MAXA(B) is only a seminorm on B but a norm on B/J . Moreover, for b in B, bMAXA(B) ≤ b + J B/J , where ·B/J denotes the quotient norm. Clearly, the vector space B/J equipped with this matrix norm satisﬁes the axioms to be an operator algebra. We call this the maximal operator algebra structure on B and denote this operator algebra by MAXA(B). Note that MAXA(B) = MAXA(B/J ) completely isometrically for J as above. An example is in order. Let B be the space of power series with summable coefﬁcients, ∞ ∞ ∞ ∞ n n an z : |an | < +∞ with an z = |a |. B= n=0 n=0 n n=0 n=0 It is well known that B, equipped with the usual Cauchy product of power series, is a Banach algebra. In fact, B can be identiﬁed with 1 (Z+ ) equipped with the convolution product. It is easily seen that π ∈ Rep(B) if and only if π (z) is a contraction. Hence, J = (0) in this case, and by von Neumann’s inequality, ∞ ∞ ∞ n n an z = an z ≤ |an |, n=0 n=0 n=0 MAXA(B)

∞

where the middle member is the supremum over the disk. Thus, MAXA(B), in this case, can be identiﬁed, completely isometrically, with the norm-dense subalgebra of A(D) consisting of power series with summable coefﬁcients. The above deﬁnition of MAXA(B) is extrinsic, and we wish to describe it intrinsically. Theorem 18.9. Let B be a normed algebra with an identity of norm 1, and let (bi j ) be in Mm,n (B). Then (bi j )MAXA(B) = inf{C1 · · · C }, where the inﬁmum is taken over all ways to factor (bi j ) = C1 D1 · · · C−1 D−1 C

270

Chapter 18. Universal Operator Algebras and Factorization

as a product of arbitrarily many matrices over B of arbitrary size, with C1 , . . . , C scalar matrices and D1 , . . . , D−1 diagonal matrices each of whose diagonal entries is an element of B of norm less than or equal to one. Proof. The proof is similar to the proof of Theorem 18.1. If we denote the quantity on the right hand side of the above equation as (bi j ) f , then it sufﬁces to prove that these satisfy the BRS conditions. We leave the details of the proof as Exercise 18.4. Corollary 18.10. Let B be a normed algebra with an identity of norm 1. Then there exists a unital isometric homomorphism from B into B(H) for some Hilbert space H if and only if for every b in B, b is given by the inﬁmum appearing in Theorem 18.9. Corollary 18.11. Let A be a unital operator algebra and let K ≥ 1. Then the following are equivalent: (i) every unital contractive homomorphism ρ from A into B(H) is completely bounded with ρcb ≤ K , (ii) the identity map id: A → MAXA(A) is completely bounded with idcb ≤ K, (iii) for every m, n and (ai j ) in Mm,n (A) with (ai j ) < 1, for some there is a factorization (ai j ) = C1 D1 · · · C−1 D−1 C with Ci scalar, C1 · · · C < K , and Di diagonal with entries from A of norm less than or equal to 1. Corollary 18.11 gives a new tool for answering questions about when contractive homomorphisms of operator algebras are completely contractive. Recall that when X is a compact subset of C, and R(X ) denotes the algebra of rational functions on X , then it is still unknown whether or not every unital contractive homomorphism of R(X ) is completely contractive. Even for X a “nice” twoholed region this is unknown. By Corollary 18.11, this question is equivalent to deciding whether or not every matrix of functions from R(X ) has a factorization as in Corollary 18.11 with K = 1. Agler [1] has proven that when X is an annulus, then every contractive homomorphism is completely contractive. However, a direct proof of the annulus case, using Corollary 18.11, is still unavailable. Such a proof might shed new light on the unknown cases. In Chapter 11, we proved that if X is a nice n-holed domain, then every contractive homomorphism ρ is completely bounded with ρcb ≤ 2n + 1, so applying Corollary 18.11 yields certain factorizations on these domains.

Chapter 18. Universal Operator Algebras and Factorization

271

In Chapter 14, we saw that for G any absorbing, absolutely convex set in Cn , n ≥ 5, there exist unital contractive homomorphisms of A(G) that are not completely contractive. Thus, there exist matrices of functions from A(G) whose factorization norm is larger than their supremum norm. That is, A(G) = MAXA(A(G)), completely isometrically. Parrott’s example shows that MAXA((P N , ·∞ )) = (P N , ·∞ ), completely isometrically, for N ≥ 3. In addition to it being unknown whether (P N , ·∞ ) and (P N , ·u ) are completely boundedly isomorphic, it is also unknown if MAXA((P N , ·∞ )) is completely boundedly isomorphic to either (P N , ·∞ ) or (P N , ·u ). We close this chapter with a factorization theorem that applies to many situations. Suppose that we are given a unital algebra A together with unital operator algebras Ai and unital homomorphisms ρi : Ai → A such that the union of the images, i ρi (Ai ), generates A as an algebra. For two examples of this situation, consider the case where A = A1 ⊗ A2 is the tensor product, with ρ1 (a1 ) = a1 ⊗ 1 and ρ2 (a2 ) = 1 ⊗ a2 , and the case where A = A1 ∗C A2 is the free product amalgamated over C, with ρ1 (a1 ) = a1 ∗ 1 and ρ2 (a2 ) = 1 ∗ a2 . We wish to endow A with an operator algebra norm that respects the algebras Ai . To this end we call a unital homomorphism π : A → B(H) admissible provided that for every i, π ◦ ρi : Ai → B(H) is completely contractive. We endow A with a matricial seminorm by deﬁning the norm of a matrix of elements from A to be the supremum over all admissible representations into all operator algebras. The set of elements of norm 0 is a two-sided ideal I in A, and the resulting matricial norm on the quotient A/I makes it into an operator algebra. We shall call this operator algebra seminorm the maximal operator algebra seminorm induced by the given family of inclusions. In the case where A = A1 ⊗ A2 is the tensor product of two operator algebras, the maximal operator algebra seminorm induced by the inclusion of A1 and A2 into A is the maximal tensor norm of [169] discussed in Chapter 12. When both algebras are C ∗ -algebras, this is, of course, the maximal C ∗ -tensor norm. In the case where A = A1 ∗C A2 , the maximal operator algebra seminorm induced by the inclusion of A1 and A2 into A is the full free product amalgamated over C introduced in Chapter 17. The above deﬁnition of this norm is extrinsic, since it needs all representations into operator algebras. Using the ideas of this chapter, we would like to characterize this norm intrinsically. Theorem 18.12. Let A j , j = 1, . . . , m, be unital operator algebras, let A be a unital algebra, and let π j : A j → A be unital homomorphisms such that the union of their images generates A algebraically. If A is in Mn (A) for some n

272

Chapter 18. Universal Operator Algebras and Factorization

and if A denotes the maximal operator algebra seminorm induced by these algebras, then A = inf{A1 · · · Am : A = πi1 (A1 ) · · · πim (Am )}, where the inﬁmum is taken over all ways to represent A as a product of images of matrices of elements from each A j in any order and involving matrices of arbitrary sizes. Proof. To prove this theorem one only needs to note that the inf on the right hand side is larger than the maximal operator algebra seminorm induced by these algebras, and then check that the right hand side deﬁnes a seminorm that satisﬁes the BRS axioms of an abstract operator algebra. The inclusion of each A j into this latter operator algebra will yield an admissible representation of A. We leave the details of this proof as Exercise 18.6.

Notes Theorems 18.1 and 18.8 and Corollaries 18.2 and 18.3 were proven in [29]. Theorems 18.6 and 18.7 were ﬁrst proven in [2]. Theorem 18.12 ﬁrst appeared in [166], where it was applied to some questions in interpolation theory. For further connections between interpolation and the abstract theory of operator algebras see [167].

Exercises 18.1 Let T be a power-bounded operator, with T n ≤ C for all n. Show that if p(z) has a factorization of the type given by Corollary 18.3, then p(T ) ≤ C . Deduce that if every polynomial had such a factorization with ≤ L, then T would be polynomially bounded with p(T ) ≤ C L p∞ for all p. Apply Theorem 10.9 to deduce the nonexistence of such an integer L. 18.2 Prove Proposition 18.4. 18.3 Let F ∈ Mm,n (Hu∞ (D N )). Prove that if Fu ≤ 1 then I − F(z)F(w)∗ can be decomposed as claimed in Theorem 18.7. 18.4 Complete the proof of Theorem 18.9. 18.5 Let A denote the algebra of Laurent polynomials. Fix R > 1, and endow A with the extrinsic operator algebra norm obtained by taking the supremum over all representations π of the Laurent polynomials determined by π (z) = T , where T is an invertible operator satisfying T ≤ R and T −1 ≤ R. Use factorization to give an intrinsic formula for this norm. 18.6 Supply the details of the proof of Theorem 18.12.

Chapter 19 Similarity and Factorization

In the last chapter we saw how the abstract characterization of operator algebras led to a number of factorization formulas for certain universal operator algebras. However, this theory was an isometric theory. In this chapter we focus on the isomorphic theory of operator algebras and applications to similarity questions. We present Pisier’s remarkable work on similarity degree and factorization degree, and Blecher’s characterization of operator algebras up to cb isomorphism. Pisier’s work shows that for an operator algebra B, every bounded homomorphism is completely bounded if and only if the type of factorization occurring in the study of MAXA(B) can be carried out with uniform control on the number of factors needed. The least such integer is the factorization degree of the algebra. Pisier’s work has a number of deep implications in the study of bounded representations of groups and in the study of Kadison’s similarity conjecture. We focus primarily on Kadison’s conjecture, that every bounded homomorphism of a C ∗ -algebra into B(H) is similar to a ∗-homomorphism. Thus, we will show that Kadison’s conjecture is equivalent to the existence of an integer d such that every C ∗ -algebra has factorization degree at most d. A pivotal role in Pisier’s work is played by the universal operator algebra of an operator space. So we begin this chapter by examining the construction of the universal algebra of a vector space in some detail. Given a vector space V , let V ⊗n denote the tensor product of V with itself n times, and set V ⊗0 = C. Given an elementary tensor x = v1 ⊗ · · · ⊗ vn ∈ V ⊗n , n = 0, and y = w1 ⊗ · · · ⊗ wm ∈ V ⊗m , m = 0, we set x y = v1 ⊗ · · · ⊗ vn ⊗ w1 ⊗ · · · ⊗ wm ∈ V ⊗(n+m) . If n = 0 and x = λ, we set λ y = (λw1 ) ⊗ · · · ⊗ wm , and if m = 0 and y = µ, we set x µ = v1 ⊗ · · · ⊗ (µvn ). This operation extends to give a bilinear pairing from V ⊗n × V ⊗m into V ⊗(n+m) , which we still denote by . 273

274

Chapter 19. Similarity and Factorization ⊕V ⊗n denote the set of ﬁnite direct sums of the tensor We let F(V ) = ∞ n=0 powers of V . For x = ⊕ xn in F(V ), we call xn the homogeneous term of degree n of x. For x = ⊕ xn and y = ⊕ yn in F(V ) we extend the deﬁnition of our bilinear pairing by setting x y = ⊕z n , where z n = nk=0 xk yn−k . This formula now extends to a bilinear map, still denoted , from F(V ) × F(V ) to F(V ). It is easily seen that F(V ) equipped with its usual vector space structure and the product becomes a unital algebra. The algebra F(V ) is called the Fock algebra of V . It has the following important universal property. If A is any unital algebra and ϕ: V → A is any linear map, then there exists a unique extension of ϕ to a unital homomorphism πϕ : F(V ) → A. Clearly, for x = v1 ⊗ · · · ⊗ vn an elementary tensor, πϕ (x) = ϕ(v1 ) · · · ϕ(vn ). Note that if (λϕ)(v) = λϕ(v), then π(λϕ) ( ⊕ xn ) = λn πϕ (xn ). Now if V is an operator space, we wish to endow F(V ) with a family of operator algebra structures. For each c > 0 and for (xi j ) ∈ Mm,n (F(V )) we set (xi j )c = sup{(πϕ (xi j )): ϕcb ≤ c}, where the supremum is taken over all Hilbert spaces H and all ϕ: V → B(H) satisfying ϕcb ≤ c. We let OAc (V ) denote F(V ) equipped with the above operator algebra norm ·c . Alternatively, we could let Vc denote the operator space V but equipped with the norm (vi j )c = c(vi j ), so that ϕ: Vc → B(H) satisﬁes ϕcb ≤ 1 if and only if ϕ: V → B(H) satisﬁes ϕcb ≤ c. Then it is easily seen that the identity map i: V → Vc induces a completely isometric isomorphism πi : OAc (V ) → OA1 (Vc ). The above deﬁnition of OAc (V ) is extrinsic, and, following the ideas from earlier chapters, we wish to give an intrinsic characterization of OAc (V ) via a factorization formula. By using Vc , as above, it will be enough to study OA1 (V ). To this end, we wish to consider all ways to factor elements of Mm,n (F(V )) as products, using the operation , of rectangular block-diagonal matrices where each block is either a scalar matrix or a matrix over V . For example, λ ⊕ v1 ⊕ (v2 ⊗ v3 + v4 ⊗ v5 )

 1 0 λ 0 0 0 = (1, 1)  0 0 v1 v2 v4 0

0 1 0 0

   0 1 0   1 . v3  1 v5

Chapter 19. Similarity and Factorization

275

In keeping with the L ∞ condition, we wish to deﬁne the norm of such a rectangular block-diagonal matrix as the maximum of the norm of the matrix over V and the norm of the scalar matrix. Theorem 19.1. Let (xi j ) ∈ Mm,n (OA1 (V )). Then (xi j )1 = inf{B1 · · · B }, where the inﬁmum is over all and all ways to factor (xi j ) = B1 · · · B as a product in F(V ) of rectangular block-diagonal matrices where each block is either scalar or a matrix over V . Proof. Let (xi j ) f denote the inﬁmum given on the right hand side. Given such a factorization and a completely contractive map ϕ: V → B(H), it is easy ϕ ϕ ϕ to see that (πϕ (xi j )) = B1 · · · B , where Bi is obtained by replacing the scalar part of Bi by scalar multiples of IH and each element v of V by ϕ(v). Since ϕ Bi ≤ Bi , it follows that (xi j )1 ≤ (xi j ) f . To complete the proof it sufﬁces to prove that F(V ) equipped with the family of matrix norms · f satisﬁes the BRS axioms. For then we can represent (F(V ), · f ) completely isometrically as an algebra of operators on B(H) via ρ: F(V ) → B(H), and let ϕ: V → B(H) be deﬁned by ϕ = ρ|V . Then πϕ = ρ and (xi j ) f = (πϕ (xi j )) ≤ (xi j )1 . The proof that (F(V ), · f ) satisﬁes the BRS axioms is similar to the proofs in Chapter 18 and is left as an exercise (Exercise 19.1). For many applications of OAc (V ) it is useful to have further information about the norm structure. In particular, for x = ⊕ xn it is important to know the relationship between xn c and xc . Proposition 19.2. For n ≥ 0, deﬁne n : OAc (V ) → OAc (V ) via

n ⊕ x j = xn . Then n is a complete contraction. Proof. Given ϕ: V → B(H), ϕ ≤ c, let ϕθ (v) = eiθ ϕ(v). Then for x =

cb 2π −inθ 1 1 πϕθ (x) dθ . Hence, πϕ (xn ) ≤ 2π

2π⊕ x j we have πϕ (xn ) = 2πi 0 e 0 πϕθ (x)dθ ≤ xc , and it follows that n is a contraction by taking the supremum over all such ϕ. The proof that n is a complete contraction is similar.

276

Chapter 19. Similarity and Factorization

Proposition 19.3. Let n ≥ 1. If V ⊗n is endowed with the Haagerup tensor norm, then the inclusion of V ⊗n into OA1 (V ) is a complete isometry. Proof. For x ∈ V ⊗n the Haagerup tensor norm of x is deﬁned as the inﬁmum of x over certain ways to represent x using the “faux” product . Comparing these factorizations of x with Theorem 19.1, we see that the factorizations used to deﬁne the Haagerup tensor product are precisely the products of the form used in Theorem 19.1 when no scalar matrices are allowed. Thus, x1 ≤ xh . To prove the other inequality, pick unital operator algebras A1 , . . . , An and complete isometries γi : V → Ai . Let A = Ai ∗C · · · ∗C An ; then by Theorem 17.6, the map γ : V ⊗n → A deﬁned by γ (v1 ⊗ · · · ⊗ vn ) = γ1 (v1 ) · · · γn (vn ) is a complete isometry from V ⊗n endowed with the Haagerup norm into A. Deﬁne ϕ: V → Mn+1 (A) via   0 γ1 (v) 0 · · · 0 ..   .. .. .. .. . . . . .   . . .  .. .. ϕ(v) =  .. 0  .   0 · · · · · · 0 γn (v) 0 ··· ··· 0 0 Then





0 · · · 0 γ1 (v1 ) · · · γn (vn )

 0 · · · 0 πϕ (v1 ⊗ · · · ⊗ vn ) = ϕ(v1 ) · · · ϕ(vn ) =  .. .. . . 0 ··· 0

0 .. .

  , 

0

and hence xh = γ (x)h = πϕ (x) ≤ x1 for x in V ⊗n . Thus, xh = x1 for x ∈ V ⊗n . The proof for matrices over V ⊗n is identical.

To obtain the analogous result for OAc (V ) one simply needs to use the identiﬁcation of OAc (V ) with OA1 (Vc ). Since the map ϕ(x) = c−1 x is a complete isometry from V to Vc , we see that the inclusion of V ⊗n into OAc (V ) satisﬁes xc = cn xh for x in V ⊗n . The ﬁrst application of the above result is to a generalization of the BRS theorem. We wish to extend the theorem to obtain a characterization of operator algebras up to complete isomorphism, characterizing algebras either lacking a unit or having a unit of norm larger than one or having a product that is only completely bounded.

Chapter 19. Similarity and Factorization

277

We claim that all of these cases can be subsumed under a single case, namely, the case of an algebra with a unit e such that e = 1 but where we only have that (ai j )(bi j ) ≤ M(ai j )(bi j ), i.e., that the multiplication is completely bounded. For example, if e = c−1 and (ai j )(bi j ) ≤ M(ai j )(bi j ), then setting ac = ca yields a completely isomorphic norm such that ec = 1 but (ai j )(bi j )c ≤ c−1 M(ai j )c (bi j )c . For further details on the reduction to this case, see Exercises 19.2 and 19.3. Exercise 19.4 illustrates some of the further subtleties that arise when dealing with quotients of nonunital algebras. Theorem 19.4. Let A be an L ∞ -matrix-normed algebra with unit e, e = 1, such that for all n and all (ai j ), (bi j ) in Mn (A), (ai j )(bi j ) ≤ M(ai j )(bi j ). Then there exists a unital operator algebra B and a unital completely bounded √ algebra isomorphism π : B → A such that π cb π −1 cb ≤ ( M + 1)2 . Proof. Let V = A, ﬁx 0 < c < 1 to be determined later, and let ϕ: V → A be given by ϕ(v) = cv. By the universal property of F(V ) there exists a homomorphism πϕ : F(V ) → A that is onto. We wish to prove that πϕ is completely bounded as a map from OA1 (V ) to A. To this end, note that if (xi j ) ∈ M p,q (V ⊗n ) with (xi j )1 = (xi j )h < 1, then (xi j ) factors as (xi j ) = (vi1j ) (vi2j ) · · · (vinj ) with (vi j ) < 1 in M p ,q (V ) for some integers p , q . Consequently, (πϕ (xi j )) = ϕ vi1j ϕ vi2j · · · ϕ vinj = cn vi1j vi2j · · · vinj , where the latter products are in A. Hence, (πϕ (xi j )) ≤ cn M n−1 vi1j vi2j · · · vinj < cn M n−1 , and so πϕ cb ≤ cn M n when restricted to the subspace of homogeneous terms of degree n ≥ 1. Note for n = 0, πϕ cb = 1. Now, given any (xi j ) in M p,q (OA1 (V )), we may decompose it as a sum of homogeneous terms of de (n) gree n, (xi j ) = n (xi(n) j ), and by Proposition 19.2, (x i j )1 ≤ (x i j )1 . Thus, ∞ ∞ (n) n n−1 = 1+ πϕ x (πϕ (xi j )) ≤ c M (xi j ), ij

n=0

n=1

and so πϕ cb ≤ (1 − cM + c)/(1 − cM) as a map from OA1 (V ) to A. If we let B = OA1 (V )/ ker(πϕ ), then B is a unital operator algebra and the induced map π = π˙ ϕ is an algebraic isomorphism with πcb = πϕ cb . Finally, given any a ∈ A, we have πϕ (c−1 a) = a and so π −1 cb ≤ c−1 . Choosing c = M+1√ M yields the desired result.

278

Chapter 19. Similarity and Factorization

It is not clear what the best constant for πcb π −1 √cb is in the above theorem. −1 One can show that necessarily π cb π cb ≥ M, where M is the least constant bounding the norm of a product, but we have been unable to achieve this bound. Theorem 19.5 (Blecher). Let A be an algebra that is also an L ∞ -matrixnormed space. If the product on A is completely bounded, then for every > 0 there exists a Hilbert space H and a one-to-one completely bounded homomorphism π: A → B(H) such that π −1 : π(A) → A is also completely bounded with π cb π −1 cb < 1 + . Proof. Let V = A, ﬁx 0 < c < 1 to be determined later, and let ϕ: V → A be given by ϕ(v) = cv as in the proof of Theorem 19.4. Let F 0 (V ) ⊂ F(V ) denote the ideal obtained by deleting the C-summand. The canonical homomorphism πϕ , restricted to F 0 (V ), which we denote by 0 πϕ , is still onto. We let OA01 (V ) ⊆ OA1 (V ) denote F 0 (V ) equipped with the L ∞ -matrix norm that it inherits as a subspace. Computing π 0 as in the proof of Theorem 19.4, we have that πϕ0 cb ≤ ∞ n n−1 ϕ cb c = 1−cM , where M is the cb norm of the product map. If we let n=1 c M 0 J = OA1 (V )/ ker(πϕ0 ), then the induced map π = π˙ 0ϕ is an algebraic isomorc , while π −1 cb ≤ c−1 as before. phism between A and J with πcb ≤ 1−cM Since ker(πϕ0 ) is an ideal in OA1 (V ), we have that J is an ideal in the unital operator algebra B = OA1 (V )/ ker(πϕ0 ). The result now follows on choosing 1 < 1 + and representing B completely isometrically as an c so that 1−cM algebra of operators on a Hilbert space. In situations where both Theorem 19.4 and Theorem 19.5 apply, the principal difference is that Theorem 19.5 allows for representations of A such that π(A) is a subalgebra of B(H) whose unit is an idempotent of norm greater than one. These more general representations allow for the lack of dependence on M in π cb π −1 cb . This lack of dependence on M implies that π(A) must have a product whose cb norm is strictly less than 1. To illustrate how this can occur, consider B ⊆ B(H) any unital subalgebra. Then for any t > 0, Bt = is an algebra with unit ( 0I

b tb : b ∈ B ⊆ B(H ⊕ H) 0 0

tI 0

) of norm

√

1 + t 2 . This algebra has a “super”

Chapter 19. Similarity and Factorization contractive multiplication, since b1 tb1 b2 tb2 ≤ √ 1 0 0 0 0 1 + t2

279

b1 tb1 b2 tb2 0 0 0 0 .

We now wish to turn our attention to Pisier’s theory of factorization and similarity degree. This theory is a remarkable reﬁnement of Corollary 18.11 to the study of when bounded homomorphisms are completely bounded. We say that a unital operator algebra A has factorization pair (d, K ) provided that for all m, n every (ai j ) ∈ Mm,n (A) with (ai j ) < 1 has a factorization (ai j ) = C1 D1 C2 · · · Cd Dd Cd+1 where C1 , . . . , Cd+1 are scalar matrices with C1 · · · Cd+1 < K and D1 , . . . , Dd are diagonal matrices with entries from the unit ball of A. We say that A has factorization degree d provided that d is the least integer such that A has factorization pair (d, K ) for some constant K . We say that A has a similarity pair (d, K ) provided that every unital bounded homomorphism ρ: A → B(H) is completely bounded with ρcb ≤ K ρd . We say that A has similarity degree d provided that d is the least integer such that A has a similarity pair (d, K ) for some constant K . One connection between factorization pairs and similarity pairs is immediate. Proposition 19.6. Let A be a unital operator algebra. If A has factorization pair (d, K ), then A has similarity pair (d, K ). Proof. Let (ai j ) < 1. Then the factorization of (ai j ) into d + 1 scalar matrices and d diagonal matrices of norm less than or equal to 1 yields a factorization of (ρ(ai j )) into the same scalar matrices and d diagonal matrices each of norm less than or equal to ρ. Hence (ρ(ai j )) ≤ K ρd . We shall prove, conversely, that if A has similarity degree d, then A has factorization degree d. The ﬁrst step in this process is to realize the connection between the Haagerup tensor product of the operator space MAX(A) with itself and the factorization degree. To this end, let VN = MAX(A) ⊗h · · · ⊗h MAX(A) denote the tensor product of N copies of MAX(A) endowed with the Haagerup tensor norm. Also, let γ N : VN → A be the completely contractive map induced by the product, so that γ N (a1 ⊗ · · · ⊗ a N ) = a1 · · · a N . If K N denotes the kernel of γ N , then, since γ N is onto, VN /K N is linearly isomorphic to A. We let ·(N ) denote the operator space structure induced on A by this identiﬁcation.

280

Chapter 19. Similarity and Factorization

Thus, for (ai j ) ∈ Mm,n (A) we have that (ai j )(1) = (ai j )MAX(A) and (ai j )(N ) < 1 if and only if (ai j ) = (γ N (u i j )) for some (u i j ) ∈ Mm,n (VN ) with (u i j )VN < 1. The following summarizes the properties of these norms. Proposition 19.7. Let A be a unital operator algebra, and let (ai j ) ∈ Mm,n (A). Then (i) (ai j )(N ) < 1 if and only if (ai j ) can be factored as (ai j ) = C1 D1 C2 D2 · · · C N D N C N +1 where C1 , . . . , C N +1 are scalar matrices with Ci < 1 and D1 , . . . , D N are diagonal matrices with entries from the unit ball of A, (ii) for all N , (ai j )(N +1) ≤ (ai j )(N ) , (iii) (ai j )MAXA(A) = lim (ai j )(N ) . N →∞

Proof. To prove (i), write (ai j ) = (γ N (u i j )) with (u i j )VN < 1. By the deﬁnition of the Haagerup tensor norm, (u i j ) = xi1j · · · xiNj , where (xikj )MAX(A) < 1. By Theorem 14.2, (xikj ) = Rk Dk Sk , where Rk and Sk are scalar matrices with Rk < 1, Sk < 1, and Dk is a diagonal matrix whose entries are from the unit ball of A. Since γ N is the map that replaces the faux product by the actual product in A, we have that (ai j ) = R1 D1 S1 R2 D2 S2 · · · R N D N S N , and the proof of (i) is completed by taking C1 = R1 , C2 = S1 R2 , . . . , C N = S N −1 R N , C N +1 = S N . Statement (ii) follows by noting that γ N (u i j ) = γ N +1 (u i j ⊗ e), where e denotes the identity of A. Statement (iii) follows from (i) and Theorem 18.9. The following is an immediate application of Proposition 19.7 and illustrates the importance of these norms. Proposition 19.8. Let A be a unital operator algebra. Then A has factorization pair (d, K ) if and only if (ai j )(d) ≤ K (ai j ) for all m, n, and (ai j ) ∈ Mm,n (A).

Chapter 19. Similarity and Factorization

281

We are now prepared to examine algebras for which every bounded homomorphism is completely bounded. Note that if this is the case, then for every c > 1 there must exist a constant M = Mc such that if ρ ≤ c then ρcb ≤ M. For if this were not true, then one could form the direct sum of a sequence of homomorphisms ρn with ρn ≤ c and ρn cb ≥ n to obtain a single homomorphism with ρ ≤ c but ρcb = +∞. Proposition 19.9. Let A be a unital operator algebra, let c > 1, and assume that for every unital homomorphism ρ: A → B(H) with ρ ≤ c we have ρcb ≤ M. Then for every m, n and every A = (ai j ) ∈ Mm,n (A) with A < 1 , there exists an integer and elements Yk ∈ Mm,n (A), 0 ≤ k ≤ , such that M (i) A = Y0 + Y1 + · · · + Y , (ii) Y0 is a scalar matrix times the identity of A with Y0 < 1, (iii) Yk (k) < c−k for 1 ≤ k ≤ . Proof. Let MAX(A)c denote the operator space MAX(A) but with (xi j )MAX(A)c = c(xi j )MAX(A) . The identity map on A induces a completely contractive homomorphism π: OA1 (MAX(A)c ) → A. Let J denote the kernel of this homomorphism, and let OA1 (MAX(A)c )/J be the quotient operator algebra, so that π˙ : OA1 (MAX(A)c )/J → A is one-to-one and onto. Let ρ = π˙ −1 : A → OA1 (MAX(A)c )/J be the inverse of this map. Since ρ(a) = a + J ≤ ac = ca, we have that ρ = c. Thus, since OA1 (MAX(A)c )/J is an abstract unital operator algebra, it must be the case that ρcb ≤ M. Consequently, if (ai j ) < M1 in Mm,n (A), then there exists an element X = (xi j ) in the open unit ball of OA1 (MAX(A)c ) with π (xi j ) = ai j . Decomposing X = X 0 + · · · + X as a sum of homogeneous terms, X k ∈ Mm,n (MAX(A)⊗k c ), we have by Proposition 19.2 that X k OA1 (MAX(A)c ) ≤ X OA1 (MAX(A)c ) < 1. Applying Proposition 19.3, we have that the norm of the element X k in OA1 (MAX(A)c ) is the same as the norm of X k in the k-fold Haagerup tensor product of MAX(A)c with itself. Since each factor is multiplied by c we see that X k Vk = c−k X k OA1 (MAX(A)c ) < c−k . Thus, if we let Yk be the image in Mm,n (A) of X k under the product map, then by deﬁnition Yk (k) < c−k for 1 ≤ k ≤ , A = Y0 + Y1 + · · · + Y , and Y0 is as claimed.

282

Chapter 19. Similarity and Factorization

We are now in a position to prove Pisier’s main factorization theorem. The following elementary lemma is useful. Lemma 19.10. Let V be a vector space with norms · and |·|. If there exists r, 0 < r < 1, such that every v ∈ V decomposes as v = v1 + v2 with a v. |||v1 ||| ≤ av, v2 ≤ r v, then |v| ≤ 1−r

Proof. Exercise 19.5.

Theorem 19.11 (Pisier). Let A be a unital operator algebra and let c > 1. If every unital homomorphism ρ: A → B(H) with ρ ≤ c satisﬁes ρcb ≤ M, then A has factorization pair M(cd+1 − 1) d, d+1 c − cd − M for any d satisfying M < cd (c − 1). Proof. Let A ∈ Mm,n (A) with A < 1/M. Write A = Y0 + Y1 + · · · + Y as in Proposition 19.9. Then Y0 + · · · + Yd (d) ≤ Y0 (d) + Y1 (d) + · · · + Yd (d) ≤ Y0 + Y1 (1) + · · · + Yd (d) ≤ 1 + c−1 + · · · + c−d , while Yd+1 + · · · + Y ≤ c−(d+1) + · · · + c− ≤

c−(d+1) c−d = 1 − c−1 c−1

by Proposition 19.7(ii). Thus the hypotheses of Lemma 19.10 are met for |·| = ·(d) with a = M(1 + c−1 + · · · + c−d ), r = Mc−d /(c − 1). Hence, A(d) ≤ K A with a , for every m, n, and A ∈ Mm,n (A). Thus, by Proposition 19.8, A has K = 1−r factorization pair (d, K ). Corollary 19.12 (Pisier). Let A be a unital operator algebra. Then A has factorization degree d if and only if A has similarity degree d. Proof. We have already seen that if A has factorization pair (d, K ) then A has similarity pair (d, K ). So assume that A has similarity pair (d, K ). Then for any c > 1, ρ ≤ c implies ρcb ≤ K cd = M. To verify that A has factorization degree d1 we

Notes

283

need K cd < cd1 (c − 1); but c is arbitrary, and taking c > K + 1 allows us to choose d1 = d. It is interesting to note that given a similarity pair (d, K ), the value of the constant K 1 that one obtains for the factorization pair (d, K 1 ) using the above results is K 1 = K (cd+1 − 1)/[c − (K + 1)], which is always larger than K . It is not known if similarity pair (d, K ) implies factorization pair (d, K ). We now return to the topic of Kadison’s conjecture, mentioned in Chapter 9. Kadison conjectured that every bounded homomorphism of a C ∗ -algebra into the algebra of bounded operators on a Hilbert space is similar to a ∗-homomorphism. By the results in Chapter 9 we see that this is equivalent to requiring that every bounded homomorphism of a C ∗ -algebra be completely bounded. The above results lead to the following conclusions. Corollary 19.13 (Pisier). The following are equivalent: (i) the answer to Kadison’s conjecture is afﬁrmative, (ii) every bounded homomorphism of a C ∗ -algebra into an operator algebra is completely bounded, (iii) for each c > 1 there exists M > 1 such that if π is a bounded homomorphism from a C ∗ -algebra into an operator algebra with π ≤ c, then π cb ≤ M, (iv) there exists c1 > 1 and M1 > 1 such that if π is a bounded homomorphism from a C ∗ -algebra into an operator algebra with π ≤ c1 , then π cb ≤ M1 , (v) there exists an integer d and constant K such that (d, K ) is a similarity pair for every C ∗ -algebra, (vi) there exists an integer d such that every C ∗ -algebra has similarity degree less than or equal to d, (vii) there exists an integer d and constant K 1 such that (d, K 1 ) is a factorization pair for every C ∗ -algebra, (viii) there exists an integer d such that every C ∗ -algebra has factorization degree less than or equal to d. Proof. We leave the proof as Exercise 19.6.

Notes Our presentation of similarity degree and factorization degree follows Pisier’s ([184], [188], and [193]) fairly closely. Equivalent versions of many of our

284

Chapter 19. Similarity and Factorization

propositions can be found there. Although the concepts of similarity pairs and factorization pairs are not mentioned explicitly, they play a role in many of his proofs. Theorem 19.5 is due to Blecher [21], but the proof that we present here is closer to Pisier’s [193]. Blecher’s original proof introduced and used the concept of a quantum variable. Theorem 19.4 seems to be new, although its proof is essentially the same as that of Theorem 19.5. The factorization appearing in Theorem 19.1 does not seem to have been noticed before. Further results on the theory of factorization degree and similarity degree for operator algebras and groups can be found in [188], [189], [190], and [193].

Exercises 19.1 Prove that (F(V ), · f ) of Theorem 19.1 satisﬁes the BRS axioms. 19.2 Let A be an L ∞ -matrix-normed algebra such that (ai j )(bi j ) ≤ M(ai j )(bi j ). Let A1 = A ⊕ C denote the algebra obtained by adjoining a unit e to A. Show that A1 can be given an L ∞ -matrix-norm structure such that e = 1 and for (ai j ), (bi j ) in Mn (A1 ), (ai j )(bi j ) ≤ (M + 2)(ai j )(bi j ). 19.3 Let A be an L ∞ -matrix-normed algebra, let B be an operator algebra, and let ϕ: A → B be a completely bounded algebra isomorphism with completely bounded inverse. Prove that for (ai j ), (bi j ) in Mn (A), one has (ai j )(bi j ) ≤ ϕ −1 cb ϕ2cb (ai j )(bi j ). 19.4 Let I0 = { f ∈ A(D): f (0) = 0}, let 0 < r < 1, and let I0,r = { f ∈ A(D): f (0) = f (r ) = 0}. Prove that I0 /I0,r is an L ∞ -matrix-normed algebra with a completely contractive multiplication and an identity e = r −1 z + I0,r . Show that e = r −1 and that dim(I0 /I0,r ) = 1. (Thus, I0 /I0,r is a one-dimensional algebra only completely isomorphic to C.) 19.5 Prove Lemma 19.10. 19.6 Prove Corollary 19.13.

Bibliography

[1] J. Agler, Rational dilation on an annulus, Ann. of Math. 121 (1985), 537–564. [2] J. Agler, On the representation of certain holomorphic functions deﬁned on a polydisc, in Topics in Operator Theory: Ernst D. Hellinger Memorial Volume, edited by L. de Branges, I. Gohberg, and J. Rovnyak, Operator Theory and Applications, Vol. 48, Birkh¨auser-Verlag, Basel, 1990, 47–66. [3] J. Agler, Some interpolation theorems of Nevanlinna–Pick type, J. Operator Theory, to appear. [4] A.B. Aleksandrov and V. Peller, Hankel operators and similarity to a contraction, Internat. Math. Res. Notices (1996), 263–275. [5] T. Ando, On a pair of commutative contractions, Acta Sci. Math. 24 (1963), 88–90. [6] W.B. Arveson, Subalgebras of C ∗ -algebras, Acta Math. 123 (1969), 141–224. [7] W.B. Arveson, Subalgebras of C ∗ -algebras II, Acta Math. 128 (1972), 271–308. [8] W.B. Arveson, An Introduction to C ∗ -Algebras, Springer-Verlag, New York, 1976. [9] W.B. Arveson, Notes on extensions of C ∗ -algebras, Duke Math. J. 44 (1977), 329–355. [10] D. Avitzour, Free products of C ∗ -algebras, Trans. Amer. Math. Soc. 271 (1982), 423–436. [11] C. Badea and V. Paulsen, Schur multipliers and operator-valued Foguel–Hankel operators, Indiana Univ. Math. J. 50 (2001), 1509–1522. [12] B.A. Barnes, The similarity problem for representations of a B ∗ -algebra, Michigan Math. J. 22 (1975), 25–32. [13] B.A. Barnes, When is a representation of a Banach algebra Naimark related to a ∗-representation?, Paciﬁc J. Math. 72 (1977), 5–25. [14] G. Bennett, Schur multipliers, Duke Math. J. 44 (1977), 603–639. [15] C.A. Berger, A strange dilation theorem, Abstract 625–152, Notices Amer. Math. Soc. 12 (l965), 590. [16] C.A. Berger and J.G. Stampﬂi, Norm relations and skew dilations, Acta Sci. Math. 28 (1967), 191–195. [17] C.A. Berger and J.G. Stampﬂi, Mapping theorems for the numerical range, Amer. J. Math. 89 (1967), 1047–1055. [18] D.P. Blecher, Tensor products of operator spaces II, Canad. J. Math. 44 (1992), 75–90. [19] D.P. Blecher, The standard dual of an operator space, Paciﬁc J. Math. 153 (1992), 15–30.

285

286

Bibliography

[20] D.P. Blecher, Generalizing Grothendieck’s program, in Function Spaces, edited by K. Jarosz, Lecture Notes in Pure and Applied Math., Vol. 136, Marcel Dekker, 1992. [21] D.P. Blecher, A completely bounded characterization of operator algebras, Math. Ann. 303 (1995), 227–240. [22] D.P. Blecher, Modules over operator algebras, and maximal C ∗ -dilation, J. Funct. Anal. 169 (1999), 251–288. [23] D.P. Blecher, The Shilov boundary of an operator space and the characterization theorems, J. Funct. Anal. 182 (2001), 280–343. [24] D.P. Blecher, Multipliers and dual operator algebras, J. Funct. Anal. 183 (2001), 498–525. [25] D.P. Blecher, E.G. Effros, and V. Zarikian, One-sided M-ideals and multipliers in operator spaces, I, Paciﬁc J. Math., to appear. [26] D.P. Blecher and C. Le Merdy, On quotients of function algebras, and operator algebra structures on p , J. Operator Theory 34 (1995), 315–346. [27] D.P. Blecher, P.S. Muhly, and V.I. Paulsen, Categories of operator modules–Morita equivalence and projective modules, Mem. Amer. Math. Soc. 143, No. 681 (2000). [28] D.P. Blecher and V.I. Paulsen, Tensor products of operator spaces, J. Funct. Anal. 99 (1991), 262–292. [29] D.P. Blecher and V.I. Paulsen, Explicit constructions of universal operator algebras and applications to polynomial factorization, Proc. Amer. Math. Soc. 112 (1991), 839–850. [30] D.P. Blecher and V.I. Paulsen, Multipliers of operator spaces and the injective envelope, Paciﬁc J. Math. 200 (2001), 1–17. [31] D.P. Blecher, Z. J. Ruan, and A.M. Sinclair, A characterization of operator algebras, J. Funct. Anal. 89 (1990), 188–201. [32] D.P. Blecher and R.R. Smith, The dual of the Haagerup tensor product, J. London Math. Soc. 45 (1992), 126–144. [33] J. Bourgain, On the similarity problem for polynomially bounded operators on Hilbert space, Israel J. Math. 54 (1986), 227–241. [34] A. Brown and C. Pearcy, Introduction to Operator Theory I: Elements of Functional Analysis, Springer-Verlag, New York, 1977. [35] J.W. Bunce, Representations of strongly amenable C ∗ -algebras, Proc. Amer. Math. Soc. 32 (1972), 241–246. [36] J.W. Bunce, Approximating maps and a Stone–Weierstrass theorem for C ∗ -algebras, Proc. Amer. Math. Soc. 79 (1980), 559–563. [37] J.W. Bunce, The similarity problem for representations of C ∗ -algebras, Proc. Amer. Math. Soc. 81 (1981), 409–413. [38] J.F. Carlson, and D.N. Clark, Projectivity and extensions of Hilbert modules over A(D N ), Michigan Math. J. 44 (1997), 365–373. [39] J.F. Carlson and D.N. Clark, Cohomology and extensions of Hilbert modules, J. Funct. Anal. 128 (1995), 278–306. [40] J. Carlson, D.N. Clark, C. Foias, and J. Williams, Projective Hilbert A(D)-modules, New York J. Math. 1 (1994), 26–38, electronic. [41] M.D. Choi, Positive linear maps on C ∗ -algebras, Canad. J. Math. 24 (1972), 520–529. [42] M.D. Choi, A Schwarz inequality for positive linear maps on C ∗ -algebras, Illinois J. Math. 18 (1974), 565–574. [43] M.D. Choi, Completely positive linear maps on complex matrices, Lin. Alg. and Appl. 10 (1975), 285–290.

Bibliography

287

[44] M.D. Choi, A simple C*-algebra generated by two ﬁnite order unitaries, Canad. J. Math. 31 (1979), 887–890. [45] M.D. Choi, Some assorted inequalities for positive linear maps on C ∗ -algebras, J. Operator Theory 4 (1980), 271–285. [46] M.D. Choi, Positive linear maps, in Operator Algebras and Applications, edited by R.V. Kadison, Proceedings of Symposia in Pure Mathematics, Vol. 38, AMS, Providence, 1982. [47] M.D. Choi and E.G. Effros, The completely positive lifting problem for C ∗ -algebras, Ann. of Math. 104 (1976), 585–609. [48] M.D. Choi and E.G. Effros, Separable nuclear C ∗ -algebras and injectivity, Duke Math J. 43 (1976), 309–322. [49] M.D. Choi and E.G. Effros, Injectivity and operator spaces, J. Functional Anal. 24 (1977), 156–209. [50] M.D. Choi and E.G. Effros, Nuclear C ∗ -algebras and injectivity. The general case, Indiana Univ. Math. J. 26 (1977), 443–446. [51] M.D. Choi and E.G. Effros, Nuclear C ∗ -algebras and the approximation property, Amer. J. Math. 100 (1978), 61–79. [52] E. Christensen, Perturbation of operator algebras II, Indiana Univ. Math. J. 26 (1977), 891–904. [53] E. Christensen, Extensions of derivations, J. Funct. Anal. 27 (1978), 234–247. [54] E. Christensen, Extensions of derivations II, Math. Scand. 50 (1982), 111– 122. [55] E. Christensen, On non-selfadjoint representations of operator algebras, Amer. J. Math. 103 (1981), 817–834. [56] E. Christensen, Similarities of I I1 factors with property , J. Operator Theory 15 (1986), 281–288. [57] E. Christensen, E.G. Effros, and A.M. Sinclair, Completely bounded multilinear maps and C ∗ -algebraic cohomology, Invent. Math. 90 (1987), 279–296. [58] E. Christensen and A.M. Sinclair, Representations of completely bounded multilinear operators, J. Funct. Anal. 72 (1987), 151–181. [59] E. Christensen and A.M. Sinclair, A survey of completely bounded operators, Bull. London Math. Soc. 21 (1989), 417–448. [60] E. Christensen and A.M. Sinclair, On von Neumann algebras which are complemented subspaces of B(H ), J. Funct. Anal. 122 (1994), 91–102. [61] E. Christensen and A.M. Sinclair, Module mappings into von Neumann algebras and injectivity, Proc. London Math. Soc. (3) 71 (1995), 618–640. [62] E. Christensen and A.M. Sinclair, Completely bounded isomorphisms of injective von Neumann algebras, Proc. Edinburgh Math. Soc. 32 (1989), 317–327. [63] P. Chu, Finite dimensional representations of function algebras, Ph.D. Thesis, University of Houston, 1992. [64] D.N. Clark and S. Ferguson, Submodules of L 2 of the N -torus, Oper. Theory Adv. Appl. 115 (2000), 113–122. [65] W. Cohn, S. Ferguson, and R. Rochberg, Boundedness of higher Hankel forms, factorization in potential spaces and derivations, Proc. London Math. Soc. (3) 82 (2001), 110–130. [66] B. Cole and J. Wermer, Ando’s theorem and sums of squares, Indiana Univ. Math. J. 48 (1999), 767–791. [67] A. Connes, Classiﬁcation of injective factors, Ann. of Math. 104 (1976), 585–609. [68] J.B. Conway, Subnormal Operators, Pitman, Boston, 1981.

288

Bibliography

[69] M.J. Crabb and A.M. Davie, von Neumann’s inequality for Hilbert space operators, Bull. London Math. Soc. 7 (1975), 49–50. [70] A.T. Dash, Joint spectral sets, Rev. Roumaine Math. Pures et Appl. 16 (1971), 13–26. [71] K.R. Davidson and V.I. Paulsen, On polynomially bounded operators, J. Reine u. Angew. Math. 487 (1997), 153–170. [72] J. Dixmier, Les moyennes invariante dans les semi-groupes et leurs applications, Acta Sci. Math. Szeged 12 (1950), 213–227. [73] J. Dixmier, C ∗ -Algebras, North-Holland, New York, 1977. [74] P.G. Dixon, Banach algebras satisfying the non-unital von Neumann inequality, Bull. London Math. Soc. 27 (1995), 359–362. [75] P.G. Dixon, The von Neumann inequality for polynomials of degree greater than two, J. London Math. Soc. (2) 14 (1976), 369–375. [76] R.G. Douglas, Banach Algebra Techniques in Operator Theory, Academic Press, New York, 1972. [77] R.G. Douglas and V.I. Paulsen, Completely bounded maps and hypo-Dirichlet algebras, Acta Sci. Math. 50 (1986), 143–157. [78] N. Dunford and J.T. Schwartz, Linear Operators I; General Theory, Interscience, New York, 1958. [79] E.G. Effros, Aspects of non-commutative order, Notes for a lecture given at the Second US–Japan Seminar on C ∗ -Algebras and Applications to Physics (April 1977). [80] E.G. Effros, Advances in quantized functional analysis, Proceedings of the International Congress of Mathematicians, Berkeley, 1986, 906–916. [81] E.G. Effros and A. Kishimoto, Module maps and Hochschild–Johnson cohomology, Indiana Univ. Math. J. 36 (1987), 257–276. [82] E.G. Effros and Z.J. Ruan, On matricially normed spaces, Paciﬁc J. Math. 132 (1988), 243–264. [83] E.G. Effros and Z.J. Ruan, A new approach to operator spaces, Canad. Math. Bull. 34 (1991), 329–337. [84] E.G. Effros and Z.J. Ruan, On the abstract characterization of operator spaces, Proc. Amer. Math. Soc. 119 (1993), 579–584. [85] E.G. Effros and Z.J. Ruan, Self duality for the Haagerup tensor product and Hilbert space factorization, J. Funct. Anal. 100 (1991), 257–284. [86] E.G. Effros and Z.J. Ruan, Recent development in operator spaces, in Current Topics in Operator Algebras, Proceedings of the ICM-90 Satellite Conference Held in Nara (August 1990), World Scientiﬁc, River Edge, N.J., 1991, 146–164. [87] E.G. Effros and Z.J. Ruan, Representations of operator bimodules and their applications, J. Operator Theory 19 (1988), 137–157. [88] E.G. Effros and Z.J. Ruan, Operator Spaces, Oxford Univ. Press, Oxford, 2000. ¨ [89] L. Fejer, Uber trigonometrische Polynome, J. f¨ur Math. 146 (1915), 53–82. [90] S. Ferguson, Polynomially bounded operators and Ext groups, Proc. Amer. Math. Soc. 124 (1996), No. 9, 2779–2785. [91] S. Ferguson, Backward shift invariant operator ranges, J. Funct. Anal. 150 (1997), 526–543. [92] S. Ferguson, The Nehari problem for the Hardy space of the torus, J. Operator Theory 40 (1998), 309–321. [93] S. Ferguson, Higher-order Hankel forms and cohomology groups, Oper. Theory Adv. Appl., to appear.

Bibliography

289

[94] S. Ferguson and M. Lacey, Weak factorization for H 1 of the bidisk, preprint. [95] S. Ferguson and R. Rochberg, Higher-order Hilbert–Schmidt Hankel forms on the ball, preprint. [96] S. Ferguson and C. Sadosky, Characterizations of bounded mean oscillation on the polydisk in terms of Hankel operators and Carleson measures (with C. Sadosky), J. Anal. Math. 81 (2000), 239–267. [97] P.A. Fillmore, Notes on Operator Theory, Van Nostrand-Reinhold, New York, 1970. [98] S.R. Foguel, A counterexample to a problem of Sz.-Nagy, Proc. Amer. Math. Soc. 15 (1964), 788–790. [99] C. Foias, Sur certains th´eor`emes de J. von Neumann, concernant les ensembles spectreaux, Acta Sci. Math. 18 (1957), 15–20. [100] C. Foias and A.E. Frazho, The Commutant Lifting Approach to Interpolation Problems, Operator Theory: Advances and Applications, Vol. 44, Birkh¨auser, Basel, 1990. [101] C. Foias and J.P. Williams, On a class of polynomially bounded operators, unpublished preprint. [102] M. Frank and V.I. Paulsen, Injective envelopes of C ∗ -algebras as operator modules, Paciﬁc J. Math., to appear. [103] T. Gamelin, Uniform Algebras, Prentice-Hall, Englewood Cliffs, N.J., 1969. [104] D. Gaspar and A. Racz, An extension of a theorem of T. Ando, Michigan Math. J. 16 (1969), 377–380. [105] I.C. Gohberg and M.G. Krein, Introduction to the Theory of Linear Nonselfadjoint Operators, Translations of Mathematical Monographs, Vol. 18, AMS, Providence, 1969. [106] U. Haagerup, Solution of the similarity problem for cyclic representations of C ∗ -algebras, Ann. of Math. 118 (1983), 215–240. [107] U. Haagerup, Decomposition of completely bounded maps on operator algebras, unpublished preprint. [108] U. Haagerup, Injectivity and decomposition of completely bounded maps, in Operator Algebras and Their Connection with Topology and Ergodic Theory, Springer Lecture Notes in Mathematics, Vol. 1132, 1985, 170–222. [109] D.W. Hadwin, Dilations and Hahn decompositions for linear maps, Canad. J. Math. 33 (1981), 826–839. [110] P.R. Halmos, On Foguel’s answer to Nagy’s question, Proc. Amer. Math. Soc. 15 (1964), 791–793. [111] P.R. Halmos, Ten problems in Hilbert space, Bull. Amer. Math. Soc. 76 (1970), 887–933. [112] M. Hamana, Injective envelopes of operator systems, Publ. RIMS Kyoto Univ. 15 (1979), 773–785. [113] M. Hamana, Injective envelopes of dynamical systems, in Operator Algebras and Operator Theory, Pitman Research Notes 271, Longman, 69–77. [114] E. Heinz, Ein v. Neumannscher Satz u¨ ber beschrankte Operatoren im Hilbertschen Raum, G¨ottinger Nachr. (1952), 5–6. [115] D.A. Herrero, A Rota universal model for operators with multiply connected spectrum, Rev. Roum. Math. Pures et Appl. 21 (1976), 15–23. [116] D.A. Herrero, Approximation of Hilbert Space Operators, Pitman, Boston, 1982. [117] J.A.R. Holbrook, On the power bounded operators of Sz.-Nagy and Foias, Acta Sci. Math. 29 (1968), 299–310. [118] J.A.R. Holbrook, Multiplicative properties of the numerical radius in operator theory, J. Reine Angew. Math. 237 (1969), 166–174.

290

Bibliography

[119] J.A.R. Holbrook, Inequalities governing the operator radii associated with unitary p-dilations, Michigan Math. J. 18 (1971), 149–159. [120] J.A.R. Holbrook, Spectral dilations and polynomially bounded operators, Indiana Univ. Math. J. 20 (1971), 1027–1034. [121] J.A.R. Holbrook, Distortion coefﬁcients for crypto-contractions, Lin. Alg. and Appl. 18 (1977), 229–256. [122] J.A.R. Holbrook, Distortion coefﬁcients for crypto-unitary operators, Lin. Alg. and Appl. 19 (1978), 189–205. [123] T. Huruya, On compact completely bounded maps of C ∗ -algebras, Michigan Math. J. 30 (1983), 213–220. [124] T. Huruya, Linear maps between certain non-separable C ∗ -algebras, Proc. Amer. Math. Soc. 92 (1984), 193–197. [125] T. Huruya, Decompositions of linear maps into non-separable C ∗ -algebras, Publ. RIMS, Kyoto Univ. 21 (1985), 645–655. [126] T. Huruya and J. Tomiyama, Completely bounded maps of C ∗ -algebras, J. Operator Theory 10 (1983), 141–152. [127] M. Junge and G. Pisier, Bilinear forms on exact operator spaces and B(H ) ⊗ B(H ), GAFA 5 (1995), 329–363. [128] R.V. Kadison, A generalized Schwarz inequality and algebraic invariants for C ∗ -algebras, Ann. Math. 56 (1952), 494–503. [129] R.V. Kadison, On the orthogonalization of operator representations, Amer. J. Math. 77 (1955), 600–620. [130] T. Kato, Some mapping theorems for the numerical range, Proc. Japan Acad. 41 (1965), 652–655. [131] E. Kirchberg, C ∗ -nuclearity implies CPAP, Math. Nachr. 76 (1977), 203–212. [132] E. Kirchberg, The derivation and the similarity problem are equivalent, J. Operator Theory 36 (1996), 59–62. [133] R.A. Kunze and E.M. Stein, Uniformly bounded representations and harmonic analysis of the 2 × 2 real unimodular group, Amer. J. Math. 82 (1960), 1–62. [134] C. Lance, On nuclear C ∗ -algebras, J. Funct. Anal. 12 (1973), 157–176. [135] C. Lance, Tensor products and nuclear C ∗ -algebras, in Operator Algebras and Applications, edited by R.V. Kadison, Proceedings of Symposia in Pure Mathematics, Vol. 38, AMS, Providence, 1982. [136] A. Lebow, On von Neumann’s theory of spectral sets, J. Math. Anal. and Appl. 7 (1963), 64–90. [137] C. LeMerdy, Representations of a quotient of a subalgebra of B(X ), Math. Proc. Cambridge Phil. Soc. 119 (1996), 83–90. [138] R.I. Loebl, Contractive linear maps on C ∗ -algebras, Michigan Math. J. 22 (1975), 361–366. [139] R.I. Loebl, A Hahn decomposition for linear maps, Paciﬁc J. Math. 65 (1976), 119–133. [140] B. Mathes, A completely bounded view of Hilbert–Schmidt operators, Houston J. Math. 17 (1991), 404–418. [141] B. Mathes, Characterizations of row and column Hilbert space, J. London Math. Soc. (2) 50 (1994), 199–208. [142] B. Mathes and V.I. Paulsen, Operator ideals and operator spaces, Proc. Amer. Math. Soc. 123 (1995), 1763–1772. [143] M.J. McAsey and P.S. Muhly, Representations of non-self-adjoint crossed products, Proc. London Math. Soc. (3) 47 (1983), 128–144. [144] W. Mlak, Unitary dilations of contraction operators, Rozprawy Mat. 46 (1965), 1–88.

Bibliography

291

[145] W. Mlak, Unitary dilations in case of ordered groups, Ann. Polon. Math. 17 (1966), 321–328. [146] W. Mlak, Positive deﬁnite contraction valued functions, Bull. Acad. Polon. Sci. Ser. Sci. Math. Astronom. Phys. 15 (1967), 509–512. [147] W. Mlak, Absolutely continuous operator valued representations of function algebras, Bull. Acad. Polon. Sci. Ser. Sci. Math. Astronom. Phys. 17 (1969), 547–550. [148] W. Mlak, Decompositions and extensions of operator valued representations of function algebras, Acta. Sci. Math. (Szeged) 30 (1969), 181–193. [149] W. Mlak, Decompositions of operator-valued representations of function algebras, Studia Math. 36 (1970), 111–123. [150] M.A. Naimark, On a representation of additive operator set functions, C. R. (Dokl.) Acad. Sci. URSS 41 (1943), 359–361. [151] M.A. Naimark, Positive deﬁnite operator functions on a commutative group, Bull. (Izv.) Acad. Sci. URSS (Ser. Math.) 7 (1943), 237–244. [152] J. von Neumann, Eine Spektraltheorie f¨ur allgemeine Operatoren eines unit¨aren Raumes, Math. Nachr. 4 (1951), 258–281. [153] R. Nevanlinna, Ueber beschrankte Funktionen, die in gegebene Punkten vorgeschriebene Werte annehmen, Ann. Acad. Sci. Fenn., No. 732 (1929). [154] T. Oikhberg, Hilbertian operator spaces with few completely bounded maps, preprint. [155] L. Page, Bounded and compact vectorial Hankel operators, Trans. Amer. Math. Soc. 150 (1970), 529–540. [156] S.K. Parrott, Unitary dilations for commuting contractions, Paciﬁc J. Math. 34 (1970), 481–490. [157] S.K. Parrott, On a quotient norm and the Sz.-Nagy–Foias lifting theorem, J. Funct. Anal. 30 (1978), 311–328. [158] V.I. Paulsen, Completely bounded maps on C ∗ -algebras and invariant operator ranges, Proc. Amer. Math. Soc. 86 (1982), 91–96. [159] V.I. Paulsen, Every completely polynomially bounded operator is similar to a contraction, J. Funct. Anal. 55 (1984), 1–17. [160] V.I. Paulsen, Completely bounded homomorphisms of operator algebras, Proc. Amer. Math. Soc. 92 (1984), 225–228. [161] V.I. Paulsen, Completely Bounded Maps and Dilations, Pitman Research Notes in Mathematics Series, Vol. 146, Pitman Longman (Wiley), 1986. [162] V.I. Paulsen, Three tensor norms for operator spaces, in Mappings of Operator Algebras, edited by H. Araki and R.V. Kadison, Proceedings of the Japan–US Joint Seminar in Operators, Univ. of Pennsylvania, 1988, Progress in Mathematics, Vol. 84, Birkh¨auser, Boston, 1990, 251–264. [163] V.I. Paulsen, Representations of function algebras, abstract operator spaces and Banach space geometry, J. Funct. Anal. 109 (1992), 113–129. [164] V.I. Paulsen, The maximal operator space of a normed space, Proc. Edinburgh Math. Soc. 39 (1996), 309–323. [165] V.I. Paulsen, Resolutions of Hilbert modules, Rocky Mountain J. Math. 27 (1997), 271–297. [166] V.I. Paulsen, Operator algebras of idempotents, J. Funct. Anal. 181 (2001), 209–226. [167] V.I. Paulsen, Matrix-valued interpolation and hyperconvex sets, Integral Eqns. Operation Theory 41 (2001), 38–62. [168] V.I. Paulsen, G. Popescu, and D. Singh, On Bohr’s inequality, J. London Math. Soc., to appear.

292

Bibliography

[169] V.I. Paulsen and S.C. Power, Tensor products of non-self-adjoint operator algebras, Rocky Mountain J. Math. 20 (1990), 331–350. [170] V.I. Paulsen and R.R. Smith, Multilinear maps and tensor norms on operator systems, J. Funct. Anal. 73 (1987), 258–276. [171] V.I. Paulsen and C.Y. Suen, Commutant representations of completely bounded maps, J. Operator Theory 13 (1985), 87–101. [172] C. Pearcy, An elementary proof of the power inequality for the numerical radius, Michigan Math. J. 13 (1966), 289–291. [173] G.K. Pedersen, C ∗ -Algebras and Their Automorphism Groups, Academic Press, London, 1979. [174] V. Peller, Estimates of functions of power bounded operators on Hilbert space, J. Operator Theory 7 (1982), 341–372. [175] V. Peller, Estimates of functions of Hilbert space operators, similarity to a contraction and related function algebras, in Linear and Complex Analysis Problem Book, edited by V.P. Havin, S.V. Hruscev, and N.K. Nikolskii, Springer Lecture Notes 1043, 1994, 199–204. [176] G. Pisier, The operator Hilbert space OH, complex interpolation and tensor norms, Memoirs Amer. Math. Soc. 122, 585 (1996). [177] G. Pisier, Non-commutative vector valued L p -spaces and completely p-summing maps, Ast´erisque 247 (1998), 1–131. [178] G. Pisier, Sur les op´erateurs factorisables par OH, Comptes Rendus Acad. Sci. Paris S´er. I 316 (1993), 165–170. [179] G. Pisier, Factorization of Linear Operators and the Geometry of Banach Spaces, CBMS (Regional Conferences of the AMS), Vol. 60, 1986; reprinted with corrections, 1987. [180] G. Pisier, Completely bounded maps between sets of Banach space operators, Indiana Univ. Math. J. 39 (1990), 251–277. [181] G. Pisier, Exact operator spaces, presented at Colloque sur les Alg`ebres d’Op´erateurs, in Recent Advances in Operator Algebras (Orleans, 1992), Soc. Math. France, Ast´erisque 232 (1995), 159–186. [182] G. Pisier, Projections from a von Neumann algebra onto a subalgebra, Bull. Soc. Math. France 123 (1995), 139–153. [183] G. Pisier, Dvoretzky’s theorem for operator spaces and applications, Houston J. Math. 22 (1996), 399–416. [184] G. Pisier, Similarity Problems and Completely Bounded Maps, 2nd, expanded edition, Springer-Verlag Lecture Notes, Vol. 1618, 2001. [185] G. Pisier, A simple proof of a theorem of Kirchberg and related results on C ∗ -norms, J. Operator Theory 35 (1996), 317–335. [186] G. Pisier, Remarks on complemented subspaces of von Neumann algebras, Proc. Roy. Soc. Edinburgh 121A (1992), 1–4. [187] G. Pisier, Quadratic forms in unitary operators, Lin. Alg. and Appl. 267 (1997), 125–137. [188] G. Pisier, The similarity degree of an operator algebra, St. Petersburg Math. J. 10 (1999), 103–146. [189] G. Pisier, The similarity degree of an operator algebra II, Math. Z. 234 (2000), 53–81. [190] G. Pisier, Similarity problems and length, Taiwanese J. Math. 5 (2001), 1–17. [191] G. Pisier, A polynomially bounded operator on Hilbert space which is not similar to a contraction, J. Amer. Math. Soc. 10 (1997), 351–369. [192] G. Pisier, A simple proof of a theorem of Jean Bourgain, Michigan Math. J. 39 (1992), 475–484.

Bibliography

293

[193] G. Pisier, An Introduction to the Theory of Operator Spaces, Cambridge Univ. Press, to appear. [194] G. Pisier, Multipliers of the Hardy space H 1 and power bounded operators, Colloq. Math. 88 (2001), 57–73. [195] F. Pop, A.M. Sinclair, and R.R. Smith, Norming C ∗ -algebras by C ∗ -subalgebras, J. Funct. Anal. 175 (2000), 168–196. [196] S. Pott, Standard models under polynomial positivity conditions, J. Operator Theory 41 (1999), 365–389. [197] E. Ricard, Decompositions de H 1 , multiplicateurs de Schur et espaces d’op´erateurs, Ph.D. Thesis, Universit´e Pierre et Marie Carie (Paris VI) 2001. [198] F. Riesz and B. Sz.-Nagy, Functional Analysis, 2nd edition, New York, 1955. [199] J.R. Ringrose, Automatic continuity of derivations of operator algebras, J. London Math. Soc. 5 (1972), 432–438. [200] A.R. Robertson, Injective matricial Hilbert spaces, Math. Proc. Cambridge Philos. Soc. 110 (1991), 183–190. [201] A.R. Robertson and S. Wassermann, Completely bounded isomorphisms of injective operator systems, Bull. London Math. Soc. 21 (1989), 285–290. [202] G.C. Rota, On models for linear operators, Comm. Pure Appl. Math. 13 (1960), 468–472. [203] Z.J. Ruan, Subspaces of C∗ -algebras, J. Funct. Anal. 76 (1988), 217–230. [204] Z.J. Ruan, Injectivity and operator spaces, Trans. Amer. Math. 315 (1989), 89–104. [205] W. Rudin, Function Theory on Polydisks, Benjamin, New York, 1969. [206] B. Russo and H.A. Dye, A note on unitary operators in C ∗ -algebras, Duke Math. J. 33 (1966), 413–416. [207] S. Sakai, C ∗ -Algebras and W ∗ -Algebras, Springer-Verlag, New York, 1971. [208] D. Sarason, On spectral sets having connected complement, Acta Sci. Math. 26 (1965), 289–299. [209] D. Sarason, Generalized interpolation in H ∞ , Trans. Amer. Math. Soc. 127 (1969), 179–203. [210] J.J. Schaffer, On unitary dilations of contractions, Proc. Amer. Math. Soc. 6 (1955), 322. [211] A. Shields, Weighted shift-operators and analytic function theory, in Topics in Operator Theory, edited by C. Pearcy, AMS, Providence, 1974. [212] A.M. Sinclair and R.R. Smith, Hochschild Cohomology of von Neumann Algebras, LMS Lecture Note Series, Cambridge Univ. Press, Cambridge, 1995. [213] A.M. Sinclair and R.R. Smith, Hochschild cohomology for von Neumann algebras with Cartan subalgebras, Amer. J. Math 120 (1998), 1043–1057. [214] R.R. Smith, Completely bounded maps between C ∗ -algebras, J. London Math. Soc. 27 (1983), 157–166. [215] R.R. Smith, Private communication. [216] R.R. Smith, Completely contractive factorization of C ∗ -algebras, J. Funct. Anal. 64 (1985), 330–337. [217] R.R. Smith, Completely bounded module maps and the Haagerup tensor product, J. Funct. Anal. 102 (1991), 156–175. [218] R.R. Smith and J.D. Ward, Matrix ranges for Hilbert space operators, Amer. J. Math. 102 (1980), 1041–1081. [219] R.R. Smith and J.D. Ward, Locally isometric liftings from quotient C ∗ -algebras, Duke Math. J. 47 (1980), 621–631. [220] R.R. Smith and D. Williams, The decomposition property for C ∗ -algebras, J. Operator Theory 16 (1986), 51–74.

294

Bibliography

[221] W.F. Stinespring, Positive functions on C ∗ -algebras, Proc. Amer. Math. Soc. 6 (1955), 211–216. [222] E. Stormer, Positive Linear Maps of C ∗ -Algebras, Lecture Notes in Physics, Vol. 29, Springer-Verlag, Berlin, 1974, 85–106. [223] E. Stormer, Extension of positive maps into B(H ), J. Funct. Anal. 66 (1986), 235–254. [224] I. Suciu, Unitary dilations in case of partially ordered groups, Bull. Acad. Polon. Sci. Ser. Math. Astr. Phys. 15 (1967), 271–275. [225] C.-Y. Suen, Completely bounded maps on C ∗ -algebras, Proc. Amer. Math. Soc. 93 (1985), 81–87. [226] C.-Y. Suen, The unique representation of a self-adjoint bounded linear functional, Proc. Amer. Math. Soc. 95 (1985), 58–62. [227] B. Sz.-Nagy, On uniformly bounded linear transformations in Hilbert space, Acta Sci. Math. Szeged 11 (1947), 152–157. [228] B. Sz.-Nagy, Sur les contractions de l’espace de Hilbert, Acta Sci. Math. 15 (1953), 87–92. [229] B. Sz.-Nagy, Completely continuous operators with uniformly bounded iterates, Magyar Tud. Akad. Mat. Kutato Int. Kaze 4 (1959), 89–93. [230] B. Sz-Nagy, Products of operators of class Cρ , Rev. Roumaine Math. Pures Appl. 13 (1968), 897–899. [231] B. Sz.-Nagy and C. Foias, Harmonic Analysis of Operators on Hilbert Space, American Elsevier, New York, 1970. [232] T. Takasaki and J. Tomiyama, On the geometry of positive maps in matrix algebras, Math. Z. 184 (1983), 101–108. [233] M. Takesaki, Theory of Operator Algebras I, Springer-Verlag, Berlin, 1979. [234] J. Tomiyama, On the transpose map of matrix algebras, Proc. Amer. Math. Soc. 88 (1983), 635–638. [235] J. Tomiyama, Recent development of the theory of completely bounded maps between C ∗ -algebras, Publ. Res. Inst. Math. Sci. 19 (1983), 1283–1303. [236] N.Th. Varopoulos, On an inequality of von Neumann and an application of the metric theory of tensor products to operator theory, J. Funct. Anal. 16 (1974), 83–100. [237] N.Th. Varopoulos, A theorem on operator algebras, Math. Scand. 37 (1975), 173–182. [238] D. Voiculescu, Norm-limits of algebraic operators, Rev. Roumaine Math. Pures et Appl. 19 (1974), 371–378. [239] J.L. Walsh, The approximation of harmonic functions by harmonic polynomials and by harmonic rational functions, Bull. Amer. Math. Soc. 35 (1929), 499–544. [240] M.E. Walter, Algebraic structures determined by 3 by 3 matrix geometry, preprint. [241] J.P. Williams, Schwarz norms for operators, Paciﬁc J. Math 24 (1968), 181–188. [242] G. Wittstock, Ein operatorwertiger Hahn–Banach Satz, J. Funct. Anal. 40 (1981), 127–150. [243] G. Wittstock, On matrix order and convexity, in Functional Analysis: Surveys and Recent Results, Math. Studies, Vol. 90, North-Holland, Amsterdam, 1984, 175–188. [244] S.L. Woronowicz, Nonextendible positive maps, Comm. Math. Phys. 51 (1976), 243–282. [245] S.L. Woronowicz, Positive maps of low dimensional matrix algebras, Rep. Math. Phys. 10 (1976), 165–183.

Bibliography

295

[246] M. Youngson, Completely contractive projections on C ∗ -algebras, Quart. J. Math. Oxford 34 (1983), 507–511. [247] C. Zhang, Representations of operator spaces, J. Operator Theory 33 (1995), 327–351. [248] C. Zhang, Completely bounded Banach–Mazur distance, Proc. Edinburgh Math. Soc. (2) 40 (1997), 247–260. [249] C. Zhang, Representation and geometry of operator spaces, Ph.D. Thesis, Univ. of Houston, 1995. [250] W. Werner, Small K-groups for operator systems, preprint.

Index

α constant, 190 ∗ -strongly continuous, 51 ∗ -vector space, 175 admissible, 172, 271 Agler, J., 48, 69, 260, 267, 268, 270, 272 Alekansdrov, A.B., 148 Ando’s dilation theorem, 61, 69, 168, 195 Ando, T., 21, 55, 94 Archimedean, 176 Arveson’s extension theorem, 86, 206 Arveson, W.B., 10, 39, 55, 82, 87, 89, 94, 131 associativity, 235 Avitzour, D., 246, 247 Badea, C., 144 Barnes, B.A., 123 Bennett, G., 115 Berger, C., 37, 39, 55, 57 Berger–Foias–Lebow theorem, 48 Berger–Kato–Stampﬂi theorem, 37 Bernstein map, 24 bilinear, 239 bimodule, 102, 115 operator, 217 representation, 218 bimodule map, 56, 95 Blecher, D.P., 197, 204, 223, 237, 257, 272, 273, 278, 284 Blecher–Effros–Zarikian theorem, 236 Blecher–Ruan–Sinclair theorem, 225, 233, 237, 240, 252, 276 boundary representation, 222 Bourgain, J., 148 Brown, A., 133 Bunce, J., 55, 123 BW topology, 85

C ∗ -envelope, 206, 219 canonical shufﬂe, 97 CAR operators, 140 relations, 140 Carath´eodory’s completion theorem, 70 Carlson, J., 148 Cartesian decomposition, 2 category operator spaces, 206, 207 operator systems, 206, 207 Cauchy–Schwarz inequality, 205 Cayley transform, 133 Choi, M.D., 20, 35, 39, 40, 50, 55, 82, 173 Choi–Effros theorem, 94, 175, 177, 208 Choquet boundary, 222 Christensen, E., 123, 127, 131, 223, 248, 255, 257, 258 Christensen–Effros–Sinclair representation, 235 Christensen–Effros–Sinclair representation theorem, 217, 223, 234, 237, 240, 255 Chu, P., 48 Clark, D.N., 148 Cohn, W., 148 Cole–Wermer proof, 70 column Hilbert space, 138, 197, 205 operator space, 229 commutant, 164 commutant lifting Theorem, 66 complete order isomorphism, 176 completely bounded, 5 bilinear, 250 Christensen–Sinclair, 250 multilinear, 237, 255 multiplicatively, 250

297

298 completely contractive, 26 module, 217 product, 232 completely isometric, 26 completely positive, 4, 26, 176 completely positive deﬁnite function, 51 conditional expectation, 95 Connes, A., 94 Conway, J., 136 cotangent space, 194 Crabbe–Davies result, 21, 94 cross-norm, 95, 159 injective, 160 maximal, 159 minimal, 159, 160 projective, 163 D-set, 152 Dash, A.T., 173 Davidson, K., 137, 147, 148 decomposition, 152 canonical, 153 complete, 152 constant, 152, 155 derivation, 125 inner, 125, 131 dilation, 87 minimal, 88 Dirichlet algebra, 48, 152 hypo-, 80, 156 set, 152 Dixmier, J., 123, 131 Dixon, P., 226 doubly commute, 159 Douglas, R.G., 82, 158 Effros, E.G., 39, 173, 184, 204, 237, 240, 248, 255, 258 essential ideal, 223 extension, 209 essential, 212 injective, 209 maximal, 212 minimal, 212 rigid, 212 extrinsic, 261 F-map, 210 F-projection, 210 F-seminorm, 210 factorization degree, 279 pair, 279 factorization theorem, 255 faux product, 240 Fejer–Riesz lemma, 12

Index Ferguson, S., 148 Fillmore, P.A., 130 Fock algebra, 274 Foguel, S.R., 132, 135, 145 Foias, C., 20, 133, 148 Foias–Frazho text, 70 Frank, M., 207, 223, 237 free group, 191 free product, 240, 246 amalgamated, 247 Gamelin, T., 156 Gaspar–Racz result, 94 Gelfand–Naimark–Segal theorem, 1 Gohberg, I., 201 group, 54 amenable, 123, 131 Haagerup, U., 114, 115, 117, 122, 123, 131, 200, 258 Haar measure, 123 Hadwin, D., 55, 105, 131 Halmos’s conjecture, 128, 144, 148 Halmos, P.R., 128, 132, 135 Hamana’s theorem, 219 Hamana, M., 206, 221, 223, 237 Hankel, 68 Heinz, E., 20 Herrero, D., 129 Hilbert C∗ -modules, 201 Holbrook, J.A.R., 128, 131, 174 homogeneous degree, 274 Huruya, T., 115, 163, 207 implementing pair, 236 injective, 87, 95, 115, 206, 207, 209, 228 envelope, 206 injective envelope, 209, 213 intertwine, 66 intertwining dilation theorem, 66 intrinsic, 262 invariant mean, 123 invariant operator range, 125 inverse limit, 251 isometric dilation, 6 Jordan curve, 152 Jordan–von Neumann theorem, 121 Jorgensen, P., 22 Junge, M., 204 K-spectral set, 18, 128, 150, 151 complete, 128 joint, 169 Kadison’s conjecture, 123, 127, 273 Kadison, R.V., 123, 131 Kaijser–Varopoulos example, 63 Kato, T., 39

Index Kirchberg, E., 127 Kishimoto, A., 258 Korovkin, P.P., 24, 57 Krein, M.G., 22, 74, 201 Kronecker product, 30, 97 Kunze, R.A., 131 L ∞ -matrix norm, 181, 184 L ∞ -matrix seminorm, 241 Lacey, M., 148 Lance, C., 82, 166, 173 Lebow, A., 135, 145, 147 LeMerdy, C., 252 Loebl, R.I., 105 map bimodule, 106 imaginary part, 100 real part, 100 self-adjoint, 100 Markov–Kakutani theorem, 123 Mathes, B., 198, 201 matricially norming, 106, 115 matrix norm, 4, 179 normed space, 180 order, 4, 176 MAX(V), 188 MAXA(B), 269 McAsey–Muhly proposition, 92, 94 measure, operator-valued, 48, 104 bounded, 49 positive, 49 regular, 49 self-adjoint, 49 spectral, 49 MIN(V), 187 module, mapping, 37 moment Hausdorff, 24, 42, 57 trigonometric, 19, 42, 52, 57 multiplicative domain, 37, 55 multiplier, 223 algebra, 235 left, 236 norm, 236 n-positive, 26 Naimark, M.A., 50, 51, 55 Nehari–Page theorem, 68 Nevanlinna, R., 260, 265 Nevanlinna–Pick formula, 69 nice n-holed set, 153 normal dilation, 47, 48 joint, 169 minimal, 47 nuclear, 166 numerical radius, 21, 35

299

OA(S), 261 OA(V ), 274 off-diagonal technique, 98 Oikhberg, T., 201 operator Foguel, 136 Foguel–Hankel, 136 self-adjoint, 125, 133 Toeplitz, 172 operator A-system, 215 operator algebra, 92 abstract, 233 unital, 233 maximal, 269, 271 operator Hilbert space, 200 operator module bi-, 234 left, 234 right, 234 operator space, 26, 184 abstract, 184 concrete, 184 dual, 197 homogeneous, 200 maximal, 188 minimal, 187 operator system, 9 abstract, 184 concrete, 184 maximal tensor, 168 nuclear, 169 order unit, 176 Parrott’s example, 89, 96, 133, 271 Parrott, S.K., 94, 96, 133, 194 partially deﬁned, 118 partition of unity, 78 partitionable, 77 Paulsen, V.I., 82, 104, 114, 125, 131, 137, 144, 147, 148, 158, 173, 190, 201, 204, 207, 223, 226, 237, 247, 257, 272 peak point, 222 Pearcy, C., 133 Peller, V., 145, 148 Pisier, G., 115, 123, 128, 135, 142, 144, 147, 148, 184, 190, 200, 204, 223, 240, 248, 255, 258, 273, 279, 282–284 polynomially bounded, 128, 132, 135 completely, 128 Pop, F., 115 Popescu, G., 226 positive element, 1 formally, 19 map, 9 operator, 1 positive deﬁnite, 265 positive deﬁnite function, 52

300 Pott, S., 173 power-bounded, 132, 135, 272 Power, S.C., 173 quadratic, 57 rank, 243 representation algebra, 88 nondegenerate, 132 representation theorem, 255 representing measure, 19 Ricard, E., 144 Riesz functional calculus, 129, 136 rigidity theorem, 212 Ringrose, J.R., 125 Rochberg, R., 148 Rota model, 129 Rota’s theorem, 130, 136 row Hilbert space, 197, 205 operator space, 229 Ruan’s theorem, 181, 184, 226 Ruan, Z.J., 175, 184, 196, 204, 223, 237, 240 Rudin, W., 193 Russo–Dye result, 15, 20 Sadosky, C., 148 Sarason, D., 70, 95 Schaeffer, J.J., 71 Schatten ideal, 201 Schur product, 29, 56, 108, 115 Schwarz inequality, 21, 27, 40 Schwarz lemma, 193 semidiscrete, 96 semiinvariant, 95 semigroup, 260 algebra, 260 homomorphism, 60 operator algebra, 261 representation, 260 Silov boundary, 221 Silov, G.E., 220 similarity degree, 279 pair, 279 simple, 221 Sinclair, A.M., 115, 223, 248, 255, 257, 258 Singh, D., 226 Smith, R.R., 39, 41, 82, 113, 115, 247, 257 spanning cone, 60 spectral complete, 88 set, 17, 48, 195 spectral radius, 130 Stampﬂi, J., 39 state, 17

Index Stein, E.M., 131 Stinespring representation, 46 generalized, 102 minimal, 46 Stinespring’s dilation theorem, 43, 131 Stinespring, W.F., 33, 39, 55 Stormer, E., 20, 82 strongly continuous, 51 Suen, C.Y., 104, 115, 131 symmetric difference, 133 symmetrically normed ideal, 201 Sz.-Nagy’s dilation theorem, 7, 47, 58, 127 Sz.-Nagy, B., 124, 131, 133, 135, 136, 174 Sz.-Nagy–Foias theorem, 39, 55, 65, 70, 167 Takesaki, M., 160, 173 tensor norm Brown, 247 factorization, 247 Haagerup, 239, 240, 254 injective, 239 projective, 239 spatial, 239 tensor product associative, 255 Haagerup, 255 spatial, 160 Toeplitz, 54, 67 matrix, 19 Tomiyama, J., 39, 41 total variation, 105 trace class, 85, 95 uniform algebra, 220 unitary dilation, 6 Varopoulos, N.Th., 21, 94, 225 Voiculescu, D., 129 von Neumann algebra, 96 von Neumann’s inequality, 7, 14, 20, 22, 58, 63, 69, 226 von Neumann, J., 22 Walsh, J.L., 80 Walter, M.E., 39 Ward, J.D., 82 Weierstrass theorem, 24 Wermer, J., 22 Williams, J.P., 148, 158 Wittstock’s decomposition theorem, 104, 131 Wittstock’s extension theorem, 99, 206, 258 Wittstock, G., 114, 115, 131, 257 Woronowicz, S.L., 50, 82 Youngson, M., 228 Zarikian, V., 237 Zhang, C., 200, 204, 223, 237